# Rethinking Claude Code Best Practices — Lydia Hallie, Anthropic Lydia Hallie, Anthropic [Talk page](https://ai.engineer/talks/vlfCjWuv4es-rethinking-claude-code-best-practices) · [Video](https://www.youtube.com/watch?v=vlfCjWuv4es) ## Rethinking Claude Code Best Practices Lydia Hallie follows a Claude Code session from manual approvals to hours of unattended work. A staging push, a test runner and a login migration show where a stronger model helps and which decisions still belong in permissions, the operating system or a test. ### From correcting one session to managing dozens [0:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=0s) Earlier in the year, a Claude Code session often stopped before the job was done. The developer was there to nudge it along, correct a decision and add another emphatic reminder to `CLAUDE.md`. Before auto mode, the choice was to approve commands by hand or accept the risk of skipping permissions. Lydia Hallie, who works on Claude Code at Anthropic, remembers doing this for one or two sessions at a time. By the time of the talk, developers at Anthropic were running dozens of sessions concurrently. A session could work for hours, check its own results and even coordinate other sessions. Hallie puts the volume in perspective with a company-wide count of about 2,500 changes a day in May and more than 30,000 a day in September. At that scale, who catches a mistake while the developer is looking at another session? Hallie rebuilds one of the earlier sessions by changing two defaults. She moves it from **Opus 4.5** to **Opus 5.5**, which needs fewer mid-session corrections because it can infer more from the request and the codebase. She also switches manual permissions to auto mode, which Anthropic began making the default in August. The model takes on more of the steering, and auto mode handles routine approvals. Two questions remain. What can an approved action affect? How will anyone know whether the finished code does the right thing? ### Keep the project knowledge Claude cannot discover [3:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=194s) An instruction that helped an older model can become a distraction. Hallie describes newer models following emphatic reminders off track, making extra tool calls and producing a worse result. Her editing question for `CLAUDE.md` is whether the instruction tells Claude something it could reasonably miss by reading the project. Suppose a service still depends on a legacy data source that has no direct reference in the code. An engineer inspecting the repository could miss that dependency. So could Claude. Keep it in `CLAUDE.md`. A description of the tech stack or directions to an obvious part of the repository can usually go. Claude Code’s prompt audit checks both `CLAUDE.md` and the project’s skills for instructions the selected model no longer needs. ### Why a staging push can stop in auto mode [4:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=252s) Auto mode uses a separate classifier for most actions beyond reading and editing files in the project. The coding model proposes an action. The classifier can allow a familiar operation, such as running tests or installing dependencies, and block a command whose safety is uncertain. Consider a request to ship a change to staging. The developer knows that deployment entails pushing to a separate deploy repository and would ordinarily approve the push. Auto mode starts with trust in the current folder and repository. The deploy repository falls outside that scope, so the push stops and the developer is back at the keyboard. Where can the team record what it knows about that destination? Auto mode’s settings accept plain English naming trusted repositories, buckets and domains, along with actions to allow or reject. Those settings go to the classifier. Naming the deploy repository gives it the missing fact about this push while keeping the scope of that trust explicit. ### Use reversibility to decide what needs a deny rule [5:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=332s) In an August study Hallie cites, about 14% of people caught a clearly dangerous command. Auto mode blocked 89% of the dangerous commands in the study. That still left roughly one in ten getting through. A team that needs to prohibit a particular tool call can put a deny rule in the Claude Code harness, the software that receives tool calls and decides whether they run. The harness checks the rule first. If the call matches, it stops there, before the classifier ever sees it. Hallie’s test is whether you could take the action back if Claude got it wrong once. You can revert a bad code change. Discarding uncommitted work may destroy the only copy. Reading a secret key can put its value into the session transcript. Those last two cases justify a rule ahead of time. Take the project’s local environment file. Auto mode would ordinarily permit a read. Adding a deny rule for that file lets the harness stop a matching request, including a direct `cat` of the file. That handles the case where the proposed command identifies the file it wants to open. ### What can an approved test process read? [7:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=436s) `npm test` is the sort of command a developer and auto mode would both approve. Once it starts, it executes project code and dependencies. A compromised dependency could try to read the same environment file or the developer’s SSH keys and send their contents to an unfamiliar host. All the developer saw at approval time was `npm test`. The dependency’s file access happens inside the running process. Where could that second read be stopped while still letting Claude test its work? With the local shell sandbox enabled, the file restriction also goes to the operating system. The test can start. When its dependency tries to open the protected file, the operating system refuses the read. Child processes inherit the restrictions, so the same limit applies if the test launches another program. Sandbox settings govern which files a command can read, which directories it can write to and which network hosts it can reach. Most of the local machine is still readable by default in Hallie’s setup, so protecting SSH keys takes its own rule. The scope also matters. The sandbox covers shell commands and their descendants. Claude’s own file operations, MCP servers and hooks still run outside it on the developer’s machine. Where does each attempt to read the same file stop? A direct read names the file before execution. In the test path, the filename becomes relevant only after the dependency runs, when the operating system can enforce the restriction. - Direct file read — The proposed command names the protected environment file. - Harness blocks the call — A matching deny rule ends the request before classification. - Run npm test — The proposed command names the test runner. - Auto mode allows the test — The test process can start. - Dependency requests the file — The attempted read occurs inside the running test. - Operating system refuses the read — The sandbox applies the file restriction to the process. - Direct file read → Harness blocks the call (Matches the deny rule) - Run npm test → Auto mode allows the test (Recognized test command) - Auto mode allows the test → Dependency requests the file (Test executes dependency code) - Dependency requests the file → Operating system refuses the read (Attempts to open the protected file) ### Give the whole session a fresh machine [9:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=566s) A cloud session gets its own machine with a fresh copy of the repository. It keeps running when the developer closes the laptop. In Hallie’s example, the local environment file, SSH keys and GitHub token were never in Git, so none arrive with the checkout. Even a tool outside the shell sandbox would find no local copy of those secrets there. If the session discovers it needs an API key, it can flag the missing credential. The developer can supply a network secret so the session can use the key without sending its value to the model. Network access has a separate setting. The defaults Hallie describes include package registries and GitHub. Reaching another host requires adding its domain to the environment. Disposing of the machine leaves a pushed commit in its remote repository and a message sent through a connector with its recipient. Hallie keeps auto mode enabled and commits the project’s deny rules to the repository. Each fresh checkout brings those rules along, so checks on outgoing actions can still catch a force push or a secret in a commit. ### Does the API migration keep existing users signed in? [10:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=647s) An isolated session can still spend hours implementing the wrong behavior. Consider a request to move login to a new API. An engineer watching each change could catch an unwanted decision as it happened. An engineer coming back hours later needs a way to tell what the implementation did to the people using the old API. What would count as migrated? Hallie gives three plausible outcomes from the same prompt. One run could move login over while leaving the old API in place. Another could switch only new logins. A third could complete the switch and log everyone out. A successful login by a new user would tell a reviewer very little about someone who was already signed in. If existing sessions must survive, make that behavior the first pass-or-fail check. The implementation that signs everyone out fails there, sending Claude back to find another way. Other requirements may be observable through a build or a screenshot. Reviewing a few well-chosen checks takes less work than reconstructing each decision from the final diff. The simplest place to specify the check is in the request. Describe what done looks like, ask Claude to break the work into tasks with their own tests, and ask it to show the actual output. That last request gives the returning reviewer something to inspect. ### Make a failed check block completion [12:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=741s) A prompt leaves it up to the model to run a check. For checks that must run, Claude Code provides hooks. A hook is a script the harness invokes at a specified event. The team chooses when it runs, which command it executes and what it checks. - **After a file edit.** Run the linter so Claude gets feedback on the edit it just made. - **When Claude tries to complete a task.** Run the tests for the code it touched. A change to authentication has to pass the authentication tests. - **When Claude tries to finish the session.** Run the whole test suite so a failure elsewhere can send Claude back to fix the change. What turns a failing test into an unfinished task? Many test tools report failure with exit code `1`. In the task-completion hook Hallie describes, passing `1` through still lets Claude Code mark the task complete. The hook script needs to translate the test failure into exit code `2`. That tells the harness to block completion and sends Claude back to fix the problem. ### Give every pull request the same checks [13:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=805s) Once one session can work unattended, a developer can start several. Each cloud session has its own machine and checkout, which saves the developer from managing local worktrees to keep concurrent edits apart. Starting each session personally and reviewing each result is a reasonable stopping point. The developer still chooses which work to start and when to accept it. The pull request gives those independent sessions a common checkpoint. Continuous integration (CI) runs the same checks on every PR, regardless of how the originating session was configured or which tests Claude ran along the way. With CI autofix enabled, Claude can see a failed check and go back to fix the change. The reviewer can spend that time on decisions the checks leave open. For selected changes, GitHub’s merge-when-ready option can merge the PR once its checks pass. Hallie’s team bases that choice on risk. Every PR first gets a Claude review from a session with fresh context. A separate group of agents double-checks each bug it reports. A small documentation or copy edit can proceed on green automated checks alone. An ordinary product change may need one human reviewer, and a risky change may need two. ### Let one session coordinate the others [15:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=910s) Who keeps track when eight sessions come back with results? Hallie finds that much context switching difficult. In Claude Code Projects, in beta at the time of the talk, the developer gives one main session the job. That session acts as an orchestrator. It splits the job into tasks, starts a separate session or *thread* for each task, tracks what comes back and verifies the results. The developer has one conversation to follow. Projects lets the developer choose the models for those two roles independently. Hallie gives **Fable** as an example for the main session, where the job is deciding what to do and judging the work that comes back. An incomplete result goes back to its thread for revision until it passes. Opus 5.5 is her example for the threads that write the code. Projects also manages context automatically, taking over decisions about when to compact or clear a session. Hallie’s suggested first experiment is one task considerably bigger than you usually give Claude. Check that CI catches the failures you care about. Configure the cloud environment’s network access, secrets and setup scripts, then start the session and see where it needs you again. In the staging example, the push waits for a decision about which repository to trust. In the login example, the session-preservation check fails and sends Claude back to revise the migration. One calls for project knowledge from the developer. The other is a correction the session can make while the developer is away. ## Key takeaways - Keep the unreferenced legacy dependency in `CLAUDE.md`. Revisit directions the model can infer from the repository when auditing instructions and skills after an upgrade. - Match the control to when the risk occurs. A deny rule stops a matching direct file read before execution. A sandbox can stop a dependency from opening that file inside an approved test run. - Choose a check that distinguishes the intended behavior from a plausible wrong result. For the login migration, exercise an existing session. To make a failing test block task completion, have the hook return exit code `2`. ## Slides ### Rethinking Claude Code Best Practices Lydia Hallie of Anthropic’s Claude Code team opens with a diagram. One orange line connects five person icons, then branches into paths marked with green checks. [0:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=0s) [Title slide crediting Lydia Hallie. An orange line connects five person icons before branching into paths marked with green checks.](https://assets.ai.engineer/sites/aiecode2025/sha256/a0e71a70140a32e7df6eef09d7348b075aa0eed9d2529aaead5f0a3716ebb2e5.webp) Rethinking Claude Code Best Practices Lydia Hallie ### Manual steering and command approval in Claude Code While Claude works on a failing auth test, the developer steers it toward a code fix and repeats the auth helper instruction in CLAUDE.md. Claude then asks permission to run git reset --hard, which would discard uncommitted changes. [1:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=61s) [A failing-test timeline shows corrections above the edits and test run. A dialog asks whether to discard uncommitted changes, with Deny, Always allow and Allow once options.](https://assets.ai.engineer/sites/aiecode2025/sha256/37690135ecaf2a1981ad0d51a30b9ab15877e3fad5aa48513ab2faa28d59bf34.webp) auth.test.ts is failing, can you fix it? Fix the code, not the test Use our auth helper CLAUDE.md IMPORTANT: always use the auth helper Read Edit ✓ Edit ✓ Bash npm test ✓ Edit ✓ Edit ✓ Bash git reset --hard Allow Claude to run this command? Discard uncommitted changes $ git reset --hard Deny Always allow Allow once ### Eight parallel Claude Code sessions In this example, Claude splits one weekly task list into eight parallel sessions. Each checks its own work. Over 6h 30m, pull requests move through review or auto-merge, and the task list shows which changes have merged. [1:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=83s) [A weekly task list branches into eight cloud sessions under a 6 h 30 m label. The diagram traces pull requests through review or auto-merge, and the task list shows four merged items.](https://assets.ai.engineer/sites/aiecode2025/sha256/1174bc761ffc6535a72e00fbebc3c0ed41e457d88b62a3e28dbc18c71127c5f0.webp) acme-web This week: the auth migration, the flaky test, dependencies, the cold start crash, docs, the billing retry, the user service rename and a11y. On it. I’ll start a session for each one. Auth migration Merged Flaky auth test Merged Dependencies Merged Cold start crash Merged and 4 more 6h 30m PR review merged flaky auth test dependencies auto-merge cold start crash auth migration docs sweep auto-merge billing retry user service a11y on settings au[illegible] ### Daily code changes merged into Anthropic’s monorepo Anthropic went from merging about 2,500 code changes per day into its monorepo in May 2026 to more than 30,000 per day in September 2026, roughly twelve times the May rate. [1:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=102s) [An arrow connects 2,500 code changes merged per day in May 2026 to 30,000+ per day in September 2026.](https://assets.ai.engineer/sites/aiecode2025/sha256/1b9162344869576f6946c084748ebcb161896335167cd8a282797e021ea3917d.webp) code changes merged per day to Anthropic’s monorepo 2,500 → 30,000+ May 2026 September 2026 Redesigning source control at Anthropic, October 2026 ### In the loop A person appears on each of the five cards. Earlier Claude Code guidance put developers in that role throughout a session: they started it, steered the work, decided which commands could run, and checked the output. [1:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=118s) [Five illustrated cards show where a person intervenes in a session by steering it, approving actions, saying no, checking the output, or starting the session.](https://assets.ai.engineer/sites/aiecode2025/sha256/1c57c2d68d36c07056f00be7f44245e8add8ee98f0fef850de5a94ff014c5812.webp) In the loop Steering it Approving it Saying no Checking it Starting it ### Out of the loop When a developer steps back from a task like approving a command or checking output, the model, the harness, or the environment has to take it over. [2:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=132s) [The five intervention cards are dimmed. An arrow points down to three labels: the model, the harness, and the environment.](https://assets.ai.engineer/sites/aiecode2025/sha256/cf79d26659828bfeae6d9f2e5b34267714044ca6d93d44d16654b4b26394d076.webp) Out of the loop Steering it Approving it Saying no Checking it Starting it The model The harness The environment ### Code quality and safe execution When several Claude Code sessions run at once, how can developers check the quality of the code each session writes? If a session runs for hours without supervision, how can they tell it’s running safely? [2:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=142s) [On the left, three green check marks accompany “Is the code high-quality?” On the right, a shield with a check mark accompanies “Is it running safely?”](https://assets.ai.engineer/sites/aiecode2025/sha256/a84df97c20f63cb86a4abc374e43f877f48fa63e715fbe51d75aeb630503c14a.webp) Is the code high-quality? Is it running safely? ### Manual guidance for a failing authentication test The user steers Claude toward a code fix using the auth helper, approves edits and npm test, and rejects git reset --hard. The model menu highlights Opus 5.5 while the session still shows Opus 4.5. [2:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=158s) [Above the timeline, guidance directs the session to fix the code and use the authentication helper. Below it, person icons mark approved edits, an approved test run, and a rejected hard reset. The model menu is open.](https://assets.ai.engineer/sites/aiecode2025/sha256/6997840dba833014112f48d39fda30d3c86b00c9906ee77a2ee891f9194e56d8.webp) auth.test.ts is failing, can you fix it? Fix the code, not the test Use our auth helper CLAUDE.md IMPORTANT: always use the auth helper Timeline: Read Edit ✓ Edit ✓ Bash npm test ✓ Edit ✓ Edit ✓ Bash git reset --hard × Edit ✓ Interface text: Fable 5.2 1 Opus 5.5 2 Sonnet 5.5 3 Haiku 5.5 4 Type / for commands Manual Opus 4.5 ### Auto mode reviews commands during the test fix Hallie switches the failing-test demo to Auto mode. It allows npm test and blocks git reset --hard without asking her to approve each command. The session then continues with another edit. [2:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=176s) [On the tool timeline, a shield with a green check marks the test command as allowed, and a shield with a pink cross marks the hard reset as blocked. Both shields are labeled “auto mode.”](https://assets.ai.engineer/sites/aiecode2025/sha256/8591b5c213b1744758f3d1e4b16cd9ac9d04a14af78f7d33aa034d313ab6cffc.webp) auth.test.ts is failing, can you fix it? Timeline: Read Edit Edit Bash npm test ✓ auto mode Edit Edit Bash git reset --hard × auto mode Edit Interface text: Type / for commands Auto Opus 5.5 ### How an auth-helper instruction can send a task off track Asked to fix auth.test.ts, Claude takes a detour through billing.ts, search.ts and settings.ts. Hallie uses the highlighted CLAUDE.md rule, “always use the auth helper,” to show how emphatic instructions can lead to extra tool calls. [3:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=191s) [A prompt asks for a fix to auth.test.ts. A connector links the highlighted auth-helper rule in CLAUDE.md to a path that arcs through billing.ts, search.ts and settings.ts.](https://assets.ai.engineer/sites/aiecode2025/sha256/4d9d35e1cdb3b1b303606cf8e11816fdb9274a47794ad333db927fade9695dc6.webp) auth.test.ts is failing, can you fix it? billing.ts search.ts settings.ts CLAUDE.md 1 Think hard and be thorough. 2 Keep going until it’s done. 3 IMPORTANT: always use the auth helper 4 We’re on Node 18. 5 src/legacy still serves our mobile apps ### Documenting a legacy dependency in CLAUDE.md The highlighted line records that the mobile apps still depend on src/legacy. Hallie recommends keeping that detail in CLAUDE.md because Claude can’t infer the dependency from the code alone. [3:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=215s) [A five-line CLAUDE.md excerpt highlights its final line in green. It says that src/legacy still serves the mobile apps.](https://assets.ai.engineer/sites/aiecode2025/sha256/06764c7e260fd10d964d3cdfdf759a5bc9a83e9d65ea3fc9c20c502d3d178097.webp) CLAUDE.md 1 Think hard and be thorough. 2 Keep going until it’s done. 3 IMPORTANT: always use the auth helper 4 We’re on Node 18. 5 src/legacy still serves our mobile apps ### How auto mode handles project edits and shell commands While fixing a failing test, Claude reads and edits project files without a check. Auto mode’s classifier allows npm test and blocks git reset --hard because it’s unsure the reset is safe. Claude then continues editing. [4:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=250s) [A timeline marks npm test with a green check and git reset --hard with a pink cross. Captions say project reads and edits proceed without a check, commands judged safe are allowed, and uncertain commands are blocked.](https://assets.ai.engineer/sites/aiecode2025/sha256/03813fc8125691b6895016320f6d4f53dd5bc68456bf857fd2cb9896ca4c0efa.webp) auth.test.ts is failing, can you fix it? Timeline: Read Edit Edit Bash npm test ✓ Edit Edit Bash git reset --hard × Edit ✓ Reads and edits in your project, no check Knows it’s safe ✓ allowed Unsure it’s safe × blocked ### Configuring auto mode’s environment and permissions Auto mode’s classifier uses plain-English settings to decide which commands to allow. The example names Acme’s trusted repos, bucket and domains, allows staging deployments, and forbids sending code to third-party APIs. [4:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=288s) [A settings.json example highlights entries for trusted repositories, cloud storage and internal domains. An allow rule permits staging deployments, and a hard-deny rule forbids sending code to third-party APIs.](https://assets.ai.engineer/sites/aiecode2025/sha256/7703d56c5aa5c3054ede1ed9acba1796b5de960c2d53fd64d9380a9b8b5f7d63.webp) ~ / .claude / settings.json { "autoMode": { "environment": [ "$defaults", "Source control: github.com/acme and all repos under it", "Trusted cloud buckets: s3://acme-staging", "Trusted internal domains: *.internal.acme.com" ], "allow": ["$defaults", "Deploying to staging is allowed"], "hard_deny": ["$defaults", "Never send code to third-party APIs"] } } Interface text: Line numbers: 1 2 3 4 5 6 7 8 9 10 11 12 ### Out of the loop The model steers the work, and auto mode handles approvals. That leaves three decisions for the user: when to say no, how to check the result, and when to start a task. [5:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=312s) [Five cards run left to right. The dimmed “Steering it” and “Approving it” cards are labeled “The model” and “Auto mode.” The brighter cards read “Saying no,” “Checking it” and “Starting it.”](https://assets.ai.engineer/sites/aiecode2025/sha256/2c59e4a6d73af5c95bdb4a3ce53f9d71092ff23eeab05eee9a8b29c1745b7af3.webp) Out of the loop Steering it The model Approving it Auto mode Saying no Checking it Starting it ### Keeping it safe and verification When Claude Code runs unattended, two questions remain. How do you keep the run safe? How do you check its work? A shield represents safety, and three green check marks along a path represent verification as the run progresses. [5:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=325s) [The “Keeping it safe” panel shows a shield with a check mark over a curved path inside a cloud-labeled box. The “Verification” panel shows three green checks along a path.](https://assets.ai.engineer/sites/aiecode2025/sha256/927da61b4b44cad22c20b6cb837e6321a2f524d12e1c0039c96dbb21bf17d6f6.webp) Keeping it safe Verification ### Keeping it safe The bright shield draws attention to safety first, while verification stays dimmed. With auto mode handling approvals, Hallie asks how to keep Claude Code safe when the model can still stray from instructions. [5:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=333s) [A bright shield and check mark sit above “Keeping it safe” on the left. The Verification panel on the right is dimmed.](https://assets.ai.engineer/sites/aiecode2025/sha256/ee6c38499b42abbc53ee524b25c53c9eb3354e19ecbad3a26a42c256dfd4d8d0.webp) Keeping it safe Verification ### People caught 14% of dangerous commands and auto mode blocked 89% In its August 2026 study with 1,053 paid testers, Anthropic swapped in one dangerous command per session. About 14% of testers spotted the command. Auto mode blocked 89% of the dangerous commands. [5:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=345s) [Two icon grids compare the reported results. Purple figures mark the 14% caught by people, and white shields mark the 89% blocked by auto mode. The caption describes 1,053 paid testers and one dangerous command per session.](https://assets.ai.engineer/sites/aiecode2025/sha256/ac6bb34f8558ddc5e72f3f6aad3f64a728db73a5a88e5d1cef72fdb79ad6bcd1.webp) 14% caught by people 89% blocked by auto mode 1,053 paid testers, one dangerous command swapped in per session. Anthropic, August 2026 ### About 1 in 10 dangerous commands still went through In a study published in August 2026, Anthropic gave 1,053 paid testers one dangerous command per session. Purple figures mark the 14% of testers who caught it. Pink shields mark the roughly 1 in 10 dangerous commands auto mode let through. [6:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=360s) [The left grid still shows 14% caught by people. On the right, pink shields among gray shields illustrate the commands that went through.](https://assets.ai.engineer/sites/aiecode2025/sha256/fff4f33b21723851ee0c307fef21bafb9df5596f04e576f0baaf832d23b5def5.webp) 14% caught by people 1 in 10 still went through 1,053 paid testers, one dangerous command swapped in per session. Anthropic, August 2026 ### A deny rule blocks a hard reset before auto mode Claude Code checks the project’s deny rules first. The Bash call matches the git reset --hard pattern, so Claude Code blocks it before the auto mode classifier sees it. [6:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=363s) [An arrow leads from a Git hard-reset tool call to a box labeled “Deny rule” and “blocked.” The auto-mode box is dimmed. Below, the matching permission rule is highlighted in the project settings.](https://assets.ai.engineer/sites/aiecode2025/sha256/ffbba41b8b017b75a3d41ffa511cd5b0df17f9143b6143dd575637f8c02ada40.webp) Bash git reset --hard tool call Deny rule blocked Auto mode never sees it .claude / settings.json 1 { 2 "permissions": { 3 "deny": ["Bash(git reset --hard *)"] 4 } 5 } ### Can I take it back? Leave edits to Claude when Git can undo a mistake. Set a project rule when one wrong action could erase uncommitted work, expose a key in the transcript, or send data outside the project. [6:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=391s) [The Yes column leaves edits to code, tests and docs, along with anything Git can revert, to the model. The No column calls for a rule when an action deletes something permanently, exposes a secret or sends something outside.](https://assets.ai.engineer/sites/aiecode2025/sha256/21044490166e1957637f087679b80f7a9294063c8426dacaebeabf1b96a24998.webp) Can I take it back? Yes leave it to the model • Edits to code, tests, docs • Anything git can revert No make it a rule • Deletes something for good • Exposes a secret • Sends something outside ### Blocking access to an environment file When Claude tries to diagnose a failing authentication test, the demo shows a single deny rule blocking both attempts to open .env: a direct file read and the shell command cat .env. [6:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=416s) [Under a request to fix a failing authentication test, a timeline marks a direct file read and a shell read with crosses. A lock marks the environment file in the project tree, and the settings panel highlights its read-denial rule.](https://assets.ai.engineer/sites/aiecode2025/sha256/0b9412fc3a8534a523a5d91a55ec374658ae9324720ab690a51253c1dc2be61a.webp) auth.test.ts is failing, can you fix it? Read .env × Bash cat .env × acme-web src package.json .env README.md .claude / settings.json { "permissions": { "deny": [ "Bash(git reset --hard *)", "Read(./.env)" ] } } Interface text: Line numbers: 1 2 3 4 5 6 7 8 ### Enabling the local sandbox Running npm test loads both the project’s tests and third-party code. With the local sandbox enabled, the operating system limits which files that code can read, which folders it can write to, and which hosts it can reach. [8:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=496s) [A settings snippet shows the sandbox enabled beside the /sandbox command. An operating-system boundary surrounds the test process, with captions for file reads, folder writes and network access.](https://assets.ai.engineer/sites/aiecode2025/sha256/1b6308dacd2407e16d6e5244c1798dbf021a230b65c51eaabbf044f9aad4b762.webp) sandbox enforced by the operating system settings.json "sandbox": { "enabled": true } /sandbox auth.test.ts is failing, can you fix it? Bash npm test What npm test runs vitest run the script in package.json src/auth/auth.test.ts your tests node_modules/some-package/index.js code you didn’t write .env which files it reads ~/.ssh which folders it writes collector.example which hosts it reaches ### How the sandbox blocks access When npm test runs a dependency, the operating system enforces sandbox rules on that code too. Here, deny rules block reads of .env and files under ~/.ssh. The sandbox also blocks the connection to the unapproved host collector.example. [8:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=517s) [Three paths from dependency code end in crosses at the sandbox boundary. Deny rules protect the environment file and SSH directory, and the destination host lacks approval.](https://assets.ai.engineer/sites/aiecode2025/sha256/2fd33291f62e2283654b40e080e6b501f0293705e167d34be1095a39ff9f0723.webp) settings.json "deny": ["Read(./.env)", "Read(~/.ssh/**)"] Bash npm test sandbox enforced by the operating system What npm test runs vitest run the script in package.json src/auth/auth.test.ts your tests node_modules/some-package/index.js code you didn’t write .env deny rule ~/.ssh deny rule collector.example host not approved ### Running tests on a fresh cloud copy of the repository The cloud session runs npm test on its own machine using a fresh copy of the repository. You can close your laptop while the tests keep running. [9:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=569s) [A cloud outline encloses a timeline marked with the npm test command. Two captions explain that the run uses a fresh repository copy and can continue with the laptop closed.](https://assets.ai.engineer/sites/aiecode2025/sha256/6947ce18d7cfbec8984cc862fdb3059b0cb510644c0dc14adc8165e6eaa84982.webp) Bash npm test a fresh copy of our repo runs with your laptop closed ### Local files and credentials absent from the cloud example The cloud session runs npm test on a fresh copy of the repository. Hallie keeps her .env file, SSH keys and GitHub token outside Git, so the test runs without them. [9:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=581s) [Below the cloud test timeline, three dimmed labels identify an environment file, an SSH directory and a GitHub token as absent.](https://assets.ai.engineer/sites/aiecode2025/sha256/02ab94f4413950180654d7b5b5632de768397df5313f0f97ede818ba01ad0024.webp) Bash npm test not in there .env ~/.ssh GitHub token ## Full transcript [0:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=0s) I had some technical difficulties during [0:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=3s) I had some technical difficulties during [0:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=3s) I had some technical difficulties during my talk this morning. So, I ran back [0:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=5s) my talk this morning. So, I ran back [0:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=5s) my talk this morning. So, I ran back home and I'm just going to record it [0:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=7s) home and I'm just going to record it [0:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=7s) home and I'm just going to record it from my computer instead. So, hi [0:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=8s) from my computer instead. So, hi [0:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=8s) from my computer instead. So, hi everyone. I'm Lydia. I work on the clock [0:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=10s) everyone. I'm Lydia. I work on the clock [0:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=10s) everyone. I'm Lydia. I work on the clock code team at Anthropic. And the models [0:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=13s) code team at Anthropic. And the models [0:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=13s) code team at Anthropic. And the models have gotten really good lately. So, we [0:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=15s) have gotten really good lately. So, we [0:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=15s) have gotten really good lately. So, we have Opus 55, which is my favorite daily [0:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=17s) have Opus 55, which is my favorite daily [0:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=17s) have Opus 55, which is my favorite daily driver. And earlier this week, we also [0:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=20s) driver. And earlier this week, we also [0:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=20s) driver. And earlier this week, we also launched Fable 52. And these models are [0:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=23s) launched Fable 52. And these models are [0:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=23s) launched Fable 52. And these models are extremely good at longunning tasks and [0:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=25s) extremely good at longunning tasks and [0:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=25s) extremely good at longunning tasks and understanding your code base. and [0:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=27s) understanding your code base. and [0:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=27s) understanding your code base. and especially Fable is also really good at [0:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=29s) especially Fable is also really good at [0:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=29s) especially Fable is also really good at verifying its own output. But as these [0:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=31s) verifying its own output. But as these [0:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=31s) verifying its own output. But as these models get better, the way that we work [0:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=33s) models get better, the way that we work [0:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=33s) models get better, the way that we work with them in cloud code also changes. So [0:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=36s) with them in cloud code also changes. So [0:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=36s) with them in cloud code also changes. So earlier this year, I don't know if you [0:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=38s) earlier this year, I don't know if you [0:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=38s) earlier this year, I don't know if you remember, but sessions kind of used to [0:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=40s) remember, but sessions kind of used to [0:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=40s) remember, but sessions kind of used to look like this. So we'd use a model like [0:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=41s) look like this. So we'd use a model like [0:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=42s) look like this. So we'd use a model like Opus 45, which honestly for its time was [0:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=44s) Opus 45, which honestly for its time was [0:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=44s) Opus 45, which honestly for its time was really good. I really enjoyed working [0:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=46s) really good. I really enjoyed working [0:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=46s) really good. I really enjoyed working with Opus45. But compared to today, it [0:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=48s) with Opus45. But compared to today, it [0:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=48s) with Opus45. But compared to today, it would just stop a lot earlier. So it [0:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=50s) would just stop a lot earlier. So it [0:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=50s) would just stop a lot earlier. So it would just run for shorter amounts of [0:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=52s) would just run for shorter amounts of [0:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=52s) would just run for shorter amounts of time. And we'd also keep correcting it. [0:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=55s) time. And we'd also keep correcting it. [0:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=55s) time. And we'd also keep correcting it. We'd keep reminding it to do things, add [0:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=57s) We'd keep reminding it to do things, add [0:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=57s) We'd keep reminding it to do things, add things like important, exclamation mark, [0:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=59s) things like important, exclamation mark, [0:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=59s) things like important, exclamation mark, all these things to our cloud MD file. [1:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=61s) all these things to our cloud MD file. [1:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=61s) all these things to our cloud MD file. And also features like auto mode didn't [1:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=63s) And also features like auto mode didn't [1:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=63s) And also features like auto mode didn't exist yet. So we were either approving [1:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=65s) exist yet. So we were either approving [1:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=65s) exist yet. So we were either approving everything manually or taking the [1:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=68s) everything manually or taking the [1:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=68s) everything manually or taking the riskier approach with dangerously skip [1:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=70s) riskier approach with dangerously skip [1:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=70s) riskier approach with dangerously skip permission. So we were still very much [1:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=73s) permission. So we were still very much [1:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=73s) permission. So we were still very much in the loop as humans. But as these [1:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=75s) in the loop as humans. But as these [1:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=76s) in the loop as humans. But as these models got better at taking on bigger [1:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=77s) models got better at taking on bigger [1:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=78s) models got better at taking on bigger tasks, our sessions also changed. So [1:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=80s) tasks, our sessions also changed. So [1:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=80s) tasks, our sessions also changed. So nowadays, at least at Anthropic, our [1:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=82s) nowadays, at least at Anthropic, our [1:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=82s) nowadays, at least at Anthropic, our sessions look more like this. So we're [1:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=83s) sessions look more like this. So we're [1:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=83s) sessions look more like this. So we're running dozens of sessions at the same [1:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=85s) running dozens of sessions at the same [1:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=85s) running dozens of sessions at the same time, running for hours, verifying their [1:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=88s) time, running for hours, verifying their [1:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=88s) time, running for hours, verifying their own results along the way, and we even [1:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=90s) own results along the way, and we even [1:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=90s) own results along the way, and we even have sessions orchestrate other [1:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=92s) have sessions orchestrate other [1:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=92s) have sessions orchestrate other sessions. So whereas we used to be in [1:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=94s) sessions. So whereas we used to be in [1:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=94s) sessions. So whereas we used to be in the loop for almost every step and [1:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=96s) the loop for almost every step and [1:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=96s) the loop for almost every step and running one or just a couple sessions at [1:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=98s) running one or just a couple sessions at [1:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=98s) running one or just a couple sessions at a time, we've now essentially become [1:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=100s) a time, we've now essentially become [1:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=100s) a time, we've now essentially become managers for dozens of sessions. We also [1:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=104s) managers for dozens of sessions. We also [1:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=104s) managers for dozens of sessions. We also get a lot more done, though. Now, you [1:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=105s) get a lot more done, though. Now, you [1:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=105s) get a lot more done, though. Now, you may have seen our blog post in October [1:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=107s) may have seen our blog post in October [1:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=107s) may have seen our blog post in October about how Enthropic went from about 2500 [1:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=110s) about how Enthropic went from about 2500 [1:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=110s) about how Enthropic went from about 2500 changes a day in May to more than 30,000 [1:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=113s) changes a day in May to more than 30,000 [1:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=113s) changes a day in May to more than 30,000 a day in September. But this also means [1:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=115s) a day in September. But this also means [1:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=115s) a day in September. But this also means that we as developers just have to [1:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=117s) that we as developers just have to [1:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=117s) that we as developers just have to rethink on what we're focused on [1:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=119s) rethink on what we're focused on [1:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=119s) rethink on what we're focused on instead. So, if you think about our best [2:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=121s) instead. So, if you think about our best [2:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=121s) instead. So, if you think about our best practices from earlier this year, it was [2:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=123s) practices from earlier this year, it was [2:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=124s) practices from earlier this year, it was mainly about steering sessions, [2:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=125s) mainly about steering sessions, [2:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=125s) mainly about steering sessions, approving commands, saying no to certain [2:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=128s) approving commands, saying no to certain [2:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=128s) approving commands, saying no to certain things or restricting what cloud could [2:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=129s) things or restricting what cloud could [2:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=130s) things or restricting what cloud could do, checking the output and also [2:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=132s) do, checking the output and also [2:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=132s) do, checking the output and also starting the sessions. So everywhere [2:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=134s) starting the sessions. So everywhere [2:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=134s) starting the sessions. So everywhere that we now step out of that loop, [2:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=136s) that we now step out of that loop, [2:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=136s) that we now step out of that loop, something else has to take over. And [2:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=138s) something else has to take over. And [2:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=138s) something else has to take over. And this is either the model, the harness, [2:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=141s) this is either the model, the harness, [2:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=141s) this is either the model, the harness, or the environment that it's running in. [2:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=143s) or the environment that it's running in. [2:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=143s) or the environment that it's running in. So if we are not in a loop to approve [2:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=145s) So if we are not in a loop to approve [2:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=145s) So if we are not in a loop to approve every command, and we're running so many [2:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=147s) every command, and we're running so many [2:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=147s) every command, and we're running so many sessions at the same time, how can we [2:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=149s) sessions at the same time, how can we [2:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=149s) sessions at the same time, how can we make sure that the code it's writing is [2:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=151s) make sure that the code it's writing is [2:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=151s) make sure that the code it's writing is still high quality? And most [2:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=153s) still high quality? And most [2:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=153s) still high quality? And most importantly, if these sessions are [2:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=154s) importantly, if these sessions are [2:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=154s) importantly, if these sessions are running for hours without our [2:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=156s) running for hours without our [2:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=156s) running for hours without our supervision, how can we make sure that [2:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=158s) supervision, how can we make sure that [2:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=158s) supervision, how can we make sure that they're running safely? So let's go [2:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=160s) they're running safely? So let's go [2:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=160s) they're running safely? So let's go through that by taking this old session [2:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=162s) through that by taking this old session [2:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=162s) through that by taking this old session and seeing what we have to change to get [2:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=164s) and seeing what we have to change to get [2:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=164s) and seeing what we have to change to get to that end result that I showed [2:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=165s) to that end result that I showed [2:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=165s) to that end result that I showed earlier. Well, first of all, let's use [2:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=167s) earlier. Well, first of all, let's use [2:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=167s) earlier. Well, first of all, let's use the actual defaults we start sessions in [2:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=169s) the actual defaults we start sessions in [2:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=169s) the actual defaults we start sessions in nowadays. So that's not really something [2:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=171s) nowadays. So that's not really something [2:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=171s) nowadays. So that's not really something you should think about. So let's update [2:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=172s) you should think about. So let's update [2:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=172s) you should think about. So let's update the model here from Opus 45 to Opus 55 [2:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=175s) the model here from Opus 45 to Opus 55 [2:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=175s) the model here from Opus 45 to Opus 55 and change the manual permission to auto [2:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=177s) and change the manual permission to auto [2:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=177s) and change the manual permission to auto mode. And Opus 55 is a lot better at [3:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=180s) mode. And Opus 55 is a lot better at [3:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=180s) mode. And Opus 55 is a lot better at understanding your intent and [3:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=181s) understanding your intent and [3:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=182s) understanding your intent and understanding your code base. So we [3:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=183s) understanding your code base. So we [3:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=183s) understanding your code base. So we really don't have to steer it as much [3:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=184s) really don't have to steer it as much [3:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=184s) really don't have to steer it as much anymore midsession. And you also no [3:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=187s) anymore midsession. And you also no [3:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=187s) anymore midsession. And you also no longer have to approve every command [3:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=188s) longer have to approve every command [3:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=188s) longer have to approve every command because we started making auto mode the [3:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=190s) because we started making auto mode the [3:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=190s) because we started making auto mode the default permission mode back in August. [3:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=192s) default permission mode back in August. [3:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=192s) default permission mode back in August. Now being better at understanding your [3:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=194s) Now being better at understanding your [3:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=194s) Now being better at understanding your intent in your codebase also means that [3:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=196s) intent in your codebase also means that [3:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=196s) intent in your codebase also means that we have to rethink the instructions we [3:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=198s) we have to rethink the instructions we [3:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=198s) we have to rethink the instructions we have in our codebase. So if you followed [3:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=201s) have in our codebase. So if you followed [3:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=201s) have in our codebase. So if you followed our best practices from earlier this [3:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=202s) our best practices from earlier this [3:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=202s) our best practices from earlier this year, you may still have things like [3:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=204s) year, you may still have things like [3:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=204s) year, you may still have things like important or never do this exclamation [3:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=206s) important or never do this exclamation [3:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=206s) important or never do this exclamation mark all these things in your cloudmd [3:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=207s) mark all these things in your cloudmd [3:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=208s) mark all these things in your cloudmd file. But with today's models, you [3:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=209s) file. But with today's models, you [3:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=209s) file. But with today's models, you really don't need that and it can [3:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=211s) really don't need that and it can [3:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=211s) really don't need that and it can actually make the model go off track. So [3:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=213s) actually make the model go off track. So [3:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=213s) actually make the model go off track. So you end up with more tool calls and [3:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=215s) you end up with more tool calls and [3:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=215s) you end up with more tool calls and overall a worse results. So nowadays [3:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=217s) overall a worse results. So nowadays [3:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=218s) overall a worse results. So nowadays your cloudmd file should only include [3:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=219s) your cloudmd file should only include [3:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=219s) your cloudmd file should only include the things that claude cannot see from [3:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=221s) the things that claude cannot see from [3:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=221s) the things that claude cannot see from your code. So for example, maybe you [3:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=223s) your code. So for example, maybe you [3:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=223s) your code. So for example, maybe you have a service that's still using some [3:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=225s) have a service that's still using some [3:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=225s) have a service that's still using some legacy source somewhere, but it's not [3:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=227s) legacy source somewhere, but it's not [3:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=227s) legacy source somewhere, but it's not directly referenced in the code. That's [3:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=229s) directly referenced in the code. That's [3:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=229s) directly referenced in the code. That's still good to keep in there. But [3:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=231s) still good to keep in there. But [3:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=231s) still good to keep in there. But everything else like what stack you're [3:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=232s) everything else like what stack you're [3:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=232s) everything else like what stack you're on or where to look, you really don't [3:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=234s) on or where to look, you really don't [3:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=234s) on or where to look, you really don't need that anymore. And we actually have [3:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=236s) need that anymore. And we actually have [3:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=236s) need that anymore. And we actually have a command for this called doctor prompt [3:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=238s) a command for this called doctor prompt [3:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=238s) a command for this called doctor prompt audit which makes it really easy to [4:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=240s) audit which makes it really easy to [4:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=240s) audit which makes it really easy to optimize your environment for a specific [4:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=242s) optimize your environment for a specific [4:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=242s) optimize your environment for a specific model. So if you're on Opus 55 here and [4:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=244s) model. So if you're on Opus 55 here and [4:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=244s) model. So if you're on Opus 55 here and you run this command, it will [4:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=246s) you run this command, it will [4:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=246s) you run this command, it will automatically check your CloudMD and [4:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=247s) automatically check your CloudMD and [4:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=247s) automatically check your CloudMD and your skills to see what the new model no [4:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=250s) your skills to see what the new model no [4:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=250s) your skills to see what the new model no longer needs. Now auto mode is a [4:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=252s) longer needs. Now auto mode is a [4:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=252s) longer needs. Now auto mode is a permission mode that runs a separate [4:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=254s) permission mode that runs a separate [4:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=254s) permission mode that runs a separate classifier on pretty much everything [4:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=256s) classifier on pretty much everything [4:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=256s) classifier on pretty much everything that isn't just reading or editing files [4:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=258s) that isn't just reading or editing files [4:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=258s) that isn't just reading or editing files in your project. So it automatically [4:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=261s) in your project. So it automatically [4:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=261s) in your project. So it automatically allows commands that it knows are safe, [4:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=263s) allows commands that it knows are safe, [4:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=263s) allows commands that it knows are safe, like running tests or installing [4:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=264s) like running tests or installing [4:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=264s) like running tests or installing dependencies, but it still blocks [4:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=267s) dependencies, but it still blocks [4:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=267s) dependencies, but it still blocks commands where it's unsure whether [4:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=268s) commands where it's unsure whether [4:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=268s) commands where it's unsure whether they're safe. And normally that's pretty [4:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=270s) they're safe. And normally that's pretty [4:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=270s) they're safe. And normally that's pretty good. Uh normally that's kind of you [4:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=272s) good. Uh normally that's kind of you [4:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=272s) good. Uh normally that's kind of you know what we want. But for example here [4:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=274s) know what we want. But for example here [4:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=274s) know what we want. But for example here if we ask it to ship something to [4:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=277s) if we ask it to ship something to [4:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=277s) if we ask it to ship something to staging we would have just you know [4:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=278s) staging we would have just you know [4:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=278s) staging we would have just you know approved the push to our deploy repo but [4:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=281s) approved the push to our deploy repo but [4:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=281s) approved the push to our deploy repo but by default auto mode only trusts this [4:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=283s) by default auto mode only trusts this [4:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=283s) by default auto mode only trusts this folder and this repo. So it would block [4:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=285s) folder and this repo. So it would block [4:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=285s) folder and this repo. So it would block that command. So we're back in the loop. [4:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=288s) that command. So we're back in the loop. [4:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=288s) that command. So we're back in the loop. Now the defaults should be good in most [4:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=290s) Now the defaults should be good in most [4:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=290s) Now the defaults should be good in most cases. But if you still notice that you [4:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=292s) cases. But if you still notice that you [4:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=292s) cases. But if you still notice that you have to approve a lot of uh commands [4:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=294s) have to approve a lot of uh commands [4:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=294s) have to approve a lot of uh commands even in auto mode. The first thing to [4:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=296s) even in auto mode. The first thing to [4:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=296s) even in auto mode. The first thing to look at is settings here because you can [4:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=298s) look at is settings here because you can [4:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=298s) look at is settings here because you can configure auto mode in your own settings [5:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=300s) configure auto mode in your own settings [5:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=300s) configure auto mode in your own settings just in plain English. And this is what [5:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=302s) just in plain English. And this is what [5:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=302s) just in plain English. And this is what gets sent to the classifier which then [5:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=303s) gets sent to the classifier which then [5:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=304s) gets sent to the classifier which then decides whether to reject or allow [5:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=305s) decides whether to reject or allow [5:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=305s) decides whether to reject or allow certain commands. So here we can just [5:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=308s) certain commands. So here we can just [5:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=308s) certain commands. So here we can just write which repos, buckets, domains we [5:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=310s) write which repos, buckets, domains we [5:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=310s) write which repos, buckets, domains we trust um and what it should always allow [5:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=312s) trust um and what it should always allow [5:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=312s) trust um and what it should always allow or never do. Okay, so honestly this was [5:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=314s) or never do. Okay, so honestly this was [5:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=314s) or never do. Okay, so honestly this was pretty easy. By just using today's [5:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=316s) pretty easy. By just using today's [5:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=316s) pretty easy. By just using today's defaults, we've already gotten rid of [5:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=318s) defaults, we've already gotten rid of [5:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=318s) defaults, we've already gotten rid of most of our in the loop stuff um because [5:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=320s) most of our in the loop stuff um because [5:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=320s) most of our in the loop stuff um because the model took over the steering and [5:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=322s) the model took over the steering and [5:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=322s) the model took over the steering and auto mode took over all the approving. [5:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=324s) auto mode took over all the approving. [5:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=324s) auto mode took over all the approving. So getting out of the loop is easy. But [5:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=326s) So getting out of the loop is easy. But [5:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=326s) So getting out of the loop is easy. But the harder question now is how do we [5:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=328s) the harder question now is how do we [5:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=328s) the harder question now is how do we make sure it's okay that we're not there [5:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=330s) make sure it's okay that we're not there [5:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=330s) make sure it's okay that we're not there anymore. So first is keeping it safe and [5:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=332s) anymore. So first is keeping it safe and [5:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=332s) anymore. So first is keeping it safe and then also verification. So let's start [5:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=335s) then also verification. So let's start [5:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=335s) then also verification. So let's start with keeping it safe because everything [5:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=336s) with keeping it safe because everything [5:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=336s) with keeping it safe because everything so far so upgrading the model and using [5:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=338s) so far so upgrading the model and using [5:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=338s) so far so upgrading the model and using auto mode is just working with a model. [5:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=342s) auto mode is just working with a model. [5:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=342s) auto mode is just working with a model. And in most cases it does what you tell [5:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=344s) And in most cases it does what you tell [5:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=344s) And in most cases it does what you tell it to do. But that's not a guarantee. So [5:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=347s) it to do. But that's not a guarantee. So [5:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=347s) it to do. But that's not a guarantee. So in a study we published back in August [5:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=349s) in a study we published back in August [5:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=349s) in a study we published back in August we tested auto mode versus real people [5:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=350s) we tested auto mode versus real people [5:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=350s) we tested auto mode versus real people in the loop. and only about 14% of [5:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=353s) in the loop. and only about 14% of [5:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=353s) in the loop. and only about 14% of people caught a clearly dangerous [5:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=355s) people caught a clearly dangerous [5:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=355s) people caught a clearly dangerous command. Auto mode blocked 89% of them. [5:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=358s) command. Auto mode blocked 89% of them. [5:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=358s) command. Auto mode blocked 89% of them. So it is a lot better at it than we [6:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=360s) So it is a lot better at it than we [6:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=360s) So it is a lot better at it than we were. But about one in 10 of those [6:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=363s) were. But about one in 10 of those [6:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=363s) were. But about one in 10 of those commands still went through. So if you [6:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=365s) commands still went through. So if you [6:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=365s) commands still went through. So if you do want the guarantee, you can only [6:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=367s) do want the guarantee, you can only [6:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=367s) do want the guarantee, you can only enforce that in the harness. And this is [6:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=369s) enforce that in the harness. And this is [6:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=369s) enforce that in the harness. And this is where some of our older best practices [6:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=371s) where some of our older best practices [6:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=371s) where some of our older best practices still matters and that's permissions. So [6:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=373s) still matters and that's permissions. So [6:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=373s) still matters and that's permissions. So for example, if you have a deny rule for [6:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=375s) for example, if you have a deny rule for [6:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=375s) for example, if you have a deny rule for a specific tool call in your project, [6:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=377s) a specific tool call in your project, [6:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=377s) a specific tool call in your project, cloud code looks at this rule before the [6:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=379s) cloud code looks at this rule before the [6:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=379s) cloud code looks at this rule before the automo classifier even looks at it. In [6:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=381s) automo classifier even looks at it. In [6:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=382s) automo classifier even looks at it. In fact, if a tool call matches a deny [6:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=384s) fact, if a tool call matches a deny [6:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=384s) fact, if a tool call matches a deny rule, the classifier never even sees it. [6:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=386s) rule, the classifier never even sees it. [6:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=386s) rule, the classifier never even sees it. So having deny rules is still the best [6:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=388s) So having deny rules is still the best [6:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=388s) So having deny rules is still the best way to make sure clot actually never [6:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=390s) way to make sure clot actually never [6:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=390s) way to make sure clot actually never does something. But then how do you know [6:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=393s) does something. But then how do you know [6:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=393s) does something. But then how do you know what needs to be a rule in your project [6:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=395s) what needs to be a rule in your project [6:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=395s) what needs to be a rule in your project versus what the model can just know [6:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=397s) versus what the model can just know [6:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=397s) versus what the model can just know because they've gotten so good at it? [6:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=398s) because they've gotten so good at it? [6:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=398s) because they've gotten so good at it? Um, I think it kind of just comes down [6:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=400s) Um, I think it kind of just comes down [6:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=400s) Um, I think it kind of just comes down to if clot gets as wrong once, can I [6:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=403s) to if clot gets as wrong once, can I [6:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=403s) to if clot gets as wrong once, can I take it back? So if it writes the wrong [6:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=405s) take it back? So if it writes the wrong [6:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=405s) take it back? So if it writes the wrong code, you can still revert that. Um, so [6:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=407s) code, you can still revert that. Um, so [6:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=407s) code, you can still revert that. Um, so you can still leave that to the model, [6:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=409s) you can still leave that to the model, [6:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=409s) you can still leave that to the model, but if it throws away or uncommitted [6:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=411s) but if it throws away or uncommitted [6:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=411s) but if it throws away or uncommitted changes or it reads a key that's now in [6:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=413s) changes or it reads a key that's now in [6:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=413s) changes or it reads a key that's now in your transcript and you don't don't want [6:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=414s) your transcript and you don't don't want [6:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=414s) your transcript and you don't don't want that, that should still be a rule. So [6:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=417s) that, that should still be a rule. So [6:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=417s) that, that should still be a rule. So for example, let's say that we just [6:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=419s) for example, let's say that we just [6:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=419s) for example, let's say that we just never want Claude reading our.end file. [7:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=421s) never want Claude reading our.end file. [7:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=421s) never want Claude reading our.end file. Now, by default, auto mode is fine with [7:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=423s) Now, by default, auto mode is fine with [7:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=423s) Now, by default, auto mode is fine with this. So it would have just approved it. [7:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=425s) this. So it would have just approved it. [7:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=425s) this. So it would have just approved it. So, so if you want to make sure that [7:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=427s) So, so if you want to make sure that [7:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=427s) So, so if you want to make sure that claude can never read this file, you [7:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=429s) claude can never read this file, you [7:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=429s) claude can never read this file, you should just add a deny rule. And then [7:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=431s) should just add a deny rule. And then [7:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=431s) should just add a deny rule. And then commands like cat cannot open this file [7:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=432s) commands like cat cannot open this file [7:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=432s) commands like cat cannot open this file either. But it's not always that [7:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=434s) either. But it's not always that [7:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=434s) either. But it's not always that straightforward because sometimes [7:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=436s) straightforward because sometimes [7:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=436s) straightforward because sometimes certain commands can have side effects. [7:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=438s) certain commands can have side effects. [7:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=438s) certain commands can have side effects. So for example, let's say that npm test [7:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=440s) So for example, let's say that npm test [7:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=440s) So for example, let's say that npm test here a very normal command. It's just a [7:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=443s) here a very normal command. It's just a [7:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=443s) here a very normal command. It's just a test. So automote runs it. That's fine. [7:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=445s) test. So automote runs it. That's fine. [7:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=445s) test. So automote runs it. That's fine. But when the script is running, it runs [7:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=447s) But when the script is running, it runs [7:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=447s) But when the script is running, it runs code that you didn't write like your [7:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=449s) code that you didn't write like your [7:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=449s) code that you didn't write like your dependencies or maybe claude runs a [7:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=451s) dependencies or maybe claude runs a [7:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=451s) dependencies or maybe claude runs a script real quick that it wrote just to [7:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=453s) script real quick that it wrote just to [7:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=453s) script real quick that it wrote just to quickly check something. So, if one of [7:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=455s) quickly check something. So, if one of [7:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=455s) quickly check something. So, if one of those scripts is compromised, it could [7:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=457s) those scripts is compromised, it could [7:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=458s) those scripts is compromised, it could just open that same end file or your SSH [7:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=460s) just open that same end file or your SSH [7:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=460s) just open that same end file or your SSH keys and send those to a server you've [7:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=462s) keys and send those to a server you've [7:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=462s) keys and send those to a server you've never heard of, right? And even if we [7:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=465s) never heard of, right? And even if we [7:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=465s) never heard of, right? And even if we were still there as humans approving [7:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=467s) were still there as humans approving [7:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=467s) were still there as humans approving everything, honestly, we would have just [7:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=468s) everything, honestly, we would have just [7:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=468s) everything, honestly, we would have just clicked allow because there's nothing to [7:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=470s) clicked allow because there's nothing to [7:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=470s) clicked allow because there's nothing to say no to. It's just part of the script [7:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=472s) say no to. It's just part of the script [7:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=472s) say no to. It's just part of the script that in this example case, npm test [7:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=475s) that in this example case, npm test [7:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=475s) that in this example case, npm test runs. And permissions only decide if a [7:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=478s) runs. And permissions only decide if a [7:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=478s) runs. And permissions only decide if a tool call runs, but they cannot limit [8:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=480s) tool call runs, but they cannot limit [8:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=480s) tool call runs, but they cannot limit what it can reach once it's running. So [8:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=483s) what it can reach once it's running. So [8:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=483s) what it can reach once it's running. So we want to make sure that these kinds of [8:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=484s) we want to make sure that these kinds of [8:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=484s) we want to make sure that these kinds of scripts can still run because this is [8:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=486s) scripts can still run because this is [8:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=486s) scripts can still run because this is just part of you know how cloud can [8:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=488s) just part of you know how cloud can [8:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=488s) just part of you know how cloud can check its own work running tests. So we [8:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=491s) check its own work running tests. So we [8:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=491s) check its own work running tests. So we cannot just block npm tests but we want [8:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=493s) cannot just block npm tests but we want [8:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=493s) cannot just block npm tests but we want to limit its side effects. So really we [8:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=495s) to limit its side effects. So really we [8:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=495s) to limit its side effects. So really we just want to limit the environment that [8:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=496s) just want to limit the environment that [8:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=496s) just want to limit the environment that it's running in. So if you're running it [8:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=499s) it's running in. So if you're running it [8:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=499s) it's running in. So if you're running it locally or on your own machine you can [8:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=501s) locally or on your own machine you can [8:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=501s) locally or on your own machine you can run commands like that in a sandbox. So [8:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=504s) run commands like that in a sandbox. So [8:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=504s) run commands like that in a sandbox. So then you have control over which files [8:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=506s) then you have control over which files [8:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=506s) then you have control over which files in a command it can read, which folders [8:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=508s) in a command it can read, which folders [8:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=508s) in a command it can read, which folders it can write to, uh which hosts it can [8:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=510s) it can write to, uh which hosts it can [8:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=510s) it can write to, uh which hosts it can reach and the operating system enforces [8:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=512s) reach and the operating system enforces [8:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=512s) reach and the operating system enforces that. Also this also applies to [8:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=515s) that. Also this also applies to [8:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=515s) that. Also this also applies to everything that command then starts. So [8:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=517s) everything that command then starts. So [8:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=517s) everything that command then starts. So any child processes. So on its own that [8:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=520s) any child processes. So on its own that [8:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=520s) any child processes. So on its own that deny rule got checked by claude before [8:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=522s) deny rule got checked by claude before [8:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=522s) deny rule got checked by claude before the command runs. But all it sees there [8:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=524s) the command runs. But all it sees there [8:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=524s) the command runs. But all it sees there is npm test. So it doesn't see which [8:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=526s) is npm test. So it doesn't see which [8:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=526s) is npm test. So it doesn't see which files that scripts ends up opening. But [8:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=528s) files that scripts ends up opening. But [8:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=528s) files that scripts ends up opening. But with the sandbox on that same rule gets [8:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=531s) with the sandbox on that same rule gets [8:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=531s) with the sandbox on that same rule gets passed on to the operating system. So [8:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=533s) passed on to the operating system. So [8:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=534s) passed on to the operating system. So now when when that dependency tries to [8:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=536s) now when when that dependency tries to [8:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=536s) now when when that dependency tries to open our end file, the operating system [8:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=537s) open our end file, the operating system [8:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=538s) open our end file, the operating system just doesn't let it. And by default, it [9:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=540s) just doesn't let it. And by default, it [9:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=540s) just doesn't let it. And by default, it can still read most of your machine. So [9:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=541s) can still read most of your machine. So [9:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=541s) can still read most of your machine. So you'd want to add another rule for your [9:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=543s) you'd want to add another rule for your [9:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=543s) you'd want to add another rule for your SSH keys if you care about that. Um, and [9:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=546s) SSH keys if you care about that. Um, and [9:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=546s) SSH keys if you care about that. Um, and it also cannot reach a host that wasn't [9:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=549s) it also cannot reach a host that wasn't [9:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=549s) it also cannot reach a host that wasn't approved. But the sandbox only covers [9:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=552s) approved. But the sandbox only covers [9:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=552s) approved. But the sandbox only covers the shell commands that Cloud runs and [9:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=553s) the shell commands that Cloud runs and [9:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=553s) the shell commands that Cloud runs and whatever those start. But Cloud's own [9:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=555s) whatever those start. But Cloud's own [9:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=555s) whatever those start. But Cloud's own files, your MCP servers, and all of your [9:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=558s) files, your MCP servers, and all of your [9:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=558s) files, your MCP servers, and all of your hooks still run outside of it. So the [9:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=560s) hooks still run outside of it. So the [9:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=560s) hooks still run outside of it. So the rest of your session still runs on your [9:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=562s) rest of your session still runs on your [9:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=562s) rest of your session still runs on your own machine and everything that's on [9:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=563s) own machine and everything that's on [9:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=563s) own machine and everything that's on there. So ideally you want this whole [9:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=566s) there. So ideally you want this whole [9:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=566s) there. So ideally you want this whole thing to run somewhere where none of [9:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=568s) thing to run somewhere where none of [9:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=568s) thing to run somewhere where none of that is even there in the first place [9:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=570s) that is even there in the first place [9:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=570s) that is even there in the first place and that is a cloud session because [9:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=572s) and that is a cloud session because [9:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=572s) and that is a cloud session because there it gets its own machine with a [9:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=574s) there it gets its own machine with a [9:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=574s) there it gets its own machine with a fresh copy of a repo and this runs even [9:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=577s) fresh copy of a repo and this runs even [9:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=577s) fresh copy of a repo and this runs even when we close our laptop which is [9:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=578s) when we close our laptop which is [9:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=578s) when we close our laptop which is definitely something you want for these [9:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=580s) definitely something you want for these [9:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=580s) definitely something you want for these longer running sessions that run for [9:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=581s) longer running sessions that run for [9:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=582s) longer running sessions that run for hours. So in this case, this [9:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=584s) hours. So in this case, this [9:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=584s) hours. So in this case, this automatically gets rid of the end file [9:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=586s) automatically gets rid of the end file [9:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=586s) automatically gets rid of the end file issue because it's not in git. Uh and [9:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=588s) issue because it's not in git. Uh and [9:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=588s) issue because it's not in git. Uh and neither are my SSH keys or or GitHub [9:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=590s) neither are my SSH keys or or GitHub [9:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=590s) neither are my SSH keys or or GitHub token. And if it still needs an API key, [9:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=593s) token. And if it still needs an API key, [9:53](https://www.youtube.com/watch?v=vlfCjWuv4es&t=593s) token. And if it still needs an API key, uh you can add that as a network secret [9:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=595s) uh you can add that as a network secret [9:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=595s) uh you can add that as a network secret where it will it will tell you like, [9:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=596s) where it will it will tell you like, [9:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=596s) where it will it will tell you like, hey, I'm missing an API key midsession. [9:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=599s) hey, I'm missing an API key midsession. [9:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=599s) hey, I'm missing an API key midsession. Um so it can use it without actually [10:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=601s) Um so it can use it without actually [10:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=601s) Um so it can use it without actually ever sending it to the model. And by [10:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=603s) ever sending it to the model. And by [10:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=603s) ever sending it to the model. And by default, also in class cloud session, [10:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=605s) default, also in class cloud session, [10:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=605s) default, also in class cloud session, network access is limited to things like [10:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=608s) network access is limited to things like [10:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=608s) network access is limited to things like package registries and GitHub. So you [10:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=610s) package registries and GitHub. So you [10:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=610s) package registries and GitHub. So you have to explicitly add those domains to [10:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=612s) have to explicitly add those domains to [10:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=612s) have to explicitly add those domains to your environment for cloud to to reach [10:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=614s) your environment for cloud to to reach [10:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=614s) your environment for cloud to to reach other hosts. [10:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=616s) other hosts. [10:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=616s) other hosts. Okay. So with a cloud session, we can [10:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=618s) Okay. So with a cloud session, we can [10:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=618s) Okay. So with a cloud session, we can run these long running sessions even [10:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=620s) run these long running sessions even [10:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=620s) run these long running sessions even with our laptop closed and it keeps us [10:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=622s) with our laptop closed and it keeps us [10:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=622s) with our laptop closed and it keeps us safe because it runs in its own [10:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=624s) safe because it runs in its own [10:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=624s) safe because it runs in its own environment. Now it's still good to [10:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=627s) environment. Now it's still good to [10:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=627s) environment. Now it's still good to remember that this machine is [10:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=628s) remember that this machine is [10:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=628s) remember that this machine is disposable, but the things that leave it [10:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=630s) disposable, but the things that leave it [10:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=630s) disposable, but the things that leave it aren't. So a commit pushed or a message [10:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=633s) aren't. So a commit pushed or a message [10:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=633s) aren't. So a commit pushed or a message sent through a connector. So we still [10:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=636s) sent through a connector. So we still [10:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=636s) sent through a connector. So we still run it in auto mode with our deny rules. [10:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=638s) run it in auto mode with our deny rules. [10:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=638s) run it in auto mode with our deny rules. Just commit it into the repo. So they [10:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=639s) Just commit it into the repo. So they [10:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=639s) Just commit it into the repo. So they come along. Uh so it it could still [10:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=641s) come along. Uh so it it could still [10:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=641s) come along. Uh so it it could still catch things like a force push or a [10:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=643s) catch things like a force push or a [10:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=643s) catch things like a force push or a secret in a commit, all that kind of [10:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=645s) secret in a commit, all that kind of [10:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=645s) secret in a commit, all that kind of stuff. [10:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=647s) stuff. [10:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=647s) stuff. And beyond just safety, um we also need [10:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=650s) And beyond just safety, um we also need [10:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=650s) And beyond just safety, um we also need to make sure that claude is actually [10:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=652s) to make sure that claude is actually [10:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=652s) to make sure that claude is actually working on the right thing. So next is [10:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=654s) working on the right thing. So next is [10:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=654s) working on the right thing. So next is verification. And the nice thing about [10:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=656s) verification. And the nice thing about [10:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=656s) verification. And the nice thing about being in a loop, uh like we kind of know [10:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=658s) being in a loop, uh like we kind of know [10:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=658s) being in a loop, uh like we kind of know knew the work was good because we'd [11:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=660s) knew the work was good because we'd [11:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=660s) knew the work was good because we'd already kind of seen every change. We'd [11:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=662s) already kind of seen every change. We'd [11:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=662s) already kind of seen every change. We'd approved everything. But now cloud works [11:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=664s) approved everything. But now cloud works [11:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=664s) approved everything. But now cloud works for hours in the background. We don't [11:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=666s) for hours in the background. We don't [11:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=666s) for hours in the background. We don't have that anymore. Um, and we're not [11:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=668s) have that anymore. Um, and we're not [11:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=668s) have that anymore. Um, and we're not really going to read every file at the [11:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=670s) really going to read every file at the [11:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=670s) really going to read every file at the end either. [11:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=672s) end either. [11:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=672s) end either. So, you know, say that we ask it to move [11:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=674s) So, you know, say that we ask it to move [11:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=674s) So, you know, say that we ask it to move our o to a new API. Well, Cloud is very [11:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=677s) our o to a new API. Well, Cloud is very [11:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=677s) our o to a new API. Well, Cloud is very good at understanding what we want, but [11:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=679s) good at understanding what we want, but [11:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=679s) good at understanding what we want, but there are still a lot of small decisions [11:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=681s) there are still a lot of small decisions [11:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=681s) there are still a lot of small decisions that it has to make along the way. So, [11:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=683s) that it has to make along the way. So, [11:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=683s) that it has to make along the way. So, with the exact same prompt, you know, [11:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=685s) with the exact same prompt, you know, [11:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=685s) with the exact same prompt, you know, one run might move it over, but leave [11:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=686s) one run might move it over, but leave [11:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=686s) one run might move it over, but leave the old API in there, or one might move [11:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=689s) the old API in there, or one might move [11:29](https://www.youtube.com/watch?v=vlfCjWuv4es&t=689s) the old API in there, or one might move it, but only for new login, and another [11:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=691s) it, but only for new login, and another [11:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=691s) it, but only for new login, and another one might move it, but log everyone out. [11:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=694s) one might move it, but log everyone out. [11:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=694s) one might move it, but log everyone out. And these all look pretty reasonable, [11:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=696s) And these all look pretty reasonable, [11:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=696s) And these all look pretty reasonable, but it really helps to set up these [11:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=698s) but it really helps to set up these [11:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=698s) but it really helps to set up these verification points along the way. So [11:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=700s) verification points along the way. So [11:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=700s) verification points along the way. So Claude keeps verifying its own results. [11:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=703s) Claude keeps verifying its own results. [11:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=703s) Claude keeps verifying its own results. And those can be anything that comes [11:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=704s) And those can be anything that comes [11:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=704s) And those can be anything that comes back as a pass or fail. So tests, a [11:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=707s) back as a pass or fail. So tests, a [11:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=707s) back as a pass or fail. So tests, a build, a screenshot, whatever you have. [11:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=709s) build, a screenshot, whatever you have. [11:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=709s) build, a screenshot, whatever you have. So now we just want to make sure that we [11:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=711s) So now we just want to make sure that we [11:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=711s) So now we just want to make sure that we review these verification points. Uh [11:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=714s) review these verification points. Uh [11:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=714s) review these verification points. Uh they're a lot shorter to review than if [11:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=715s) they're a lot shorter to review than if [11:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=715s) they're a lot shorter to review than if we just had to, you know, review the [11:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=717s) we just had to, you know, review the [11:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=717s) we just had to, you know, review the whole change at the end. So So that's [11:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=719s) whole change at the end. So So that's [11:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=719s) whole change at the end. So So that's pretty reasonable. So, if one of these [12:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=721s) pretty reasonable. So, if one of these [12:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=721s) pretty reasonable. So, if one of these runs would still log everyone out, it [12:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=723s) runs would still log everyone out, it [12:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=723s) runs would still log everyone out, it doesn't get past that first check and [12:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=724s) doesn't get past that first check and [12:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=724s) doesn't get past that first check and then it just has to find another way. [12:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=726s) then it just has to find another way. [12:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=726s) then it just has to find another way. And there's a couple ways to set this [12:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=728s) And there's a couple ways to set this [12:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=728s) And there's a couple ways to set this up. Um, the easiest one is, of course, [12:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=729s) up. Um, the easiest one is, of course, [12:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=730s) up. Um, the easiest one is, of course, just in your prompt. So, you know, you [12:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=731s) just in your prompt. So, you know, you [12:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=731s) just in your prompt. So, you know, you say what done looks like and ask it to [12:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=733s) say what done looks like and ask it to [12:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=733s) say what done looks like and ask it to split that into tasks so they have their [12:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=735s) split that into tasks so they have their [12:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=735s) split that into tasks so they have their own tests. Um, and I also like to ask it [12:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=737s) own tests. Um, and I also like to ask it [12:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=737s) own tests. Um, and I also like to ask it to show the actual output because you [12:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=739s) to show the actual output because you [12:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=739s) to show the actual output because you weren't there to see it. But again, [12:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=741s) weren't there to see it. But again, [12:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=741s) weren't there to see it. But again, prompting is not a guarantee that a [12:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=743s) prompting is not a guarantee that a [12:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=743s) prompting is not a guarantee that a model will do it. So for the ones that [12:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=745s) model will do it. So for the ones that [12:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=745s) model will do it. So for the ones that have to run guaranteed run, I would use [12:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=748s) have to run guaranteed run, I would use [12:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=748s) have to run guaranteed run, I would use hooks um which are just you know scripts [12:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=750s) hooks um which are just you know scripts [12:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=750s) hooks um which are just you know scripts that cloud code runs at a specific point [12:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=753s) that cloud code runs at a specific point [12:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=753s) that cloud code runs at a specific point and those give you a lot of control [12:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=755s) and those give you a lot of control [12:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=756s) and those give you a lot of control because you decide when they run and [12:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=758s) because you decide when they run and [12:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=758s) because you decide when they run and what they run and what they check. So [12:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=760s) what they run and what they check. So [12:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=760s) what they run and what they check. So for example every time it edits a file [12:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=762s) for example every time it edits a file [12:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=762s) for example every time it edits a file we can run our llinter every time it [12:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=763s) we can run our llinter every time it [12:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=763s) we can run our llinter every time it wants to mark a task as complete we can [12:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=765s) wants to mark a task as complete we can [12:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=766s) wants to mark a task as complete we can run the tests for what it changed. Um, [12:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=768s) run the tests for what it changed. Um, [12:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=768s) run the tests for what it changed. Um, so you know if it touched our off code [12:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=769s) so you know if it touched our off code [12:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=769s) so you know if it touched our off code or something it has to pass the off [12:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=771s) or something it has to pass the off [12:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=771s) or something it has to pass the off tests and when it wants to finish the [12:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=774s) tests and when it wants to finish the [12:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=774s) tests and when it wants to finish the whole session we can just run the entire [12:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=775s) whole session we can just run the entire [12:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=775s) whole session we can just run the entire test suite and if one of those fails it [12:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=778s) test suite and if one of those fails it [12:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=778s) test suite and if one of those fails it goes oh that's still broken it goes back [13:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=780s) goes oh that's still broken it goes back [13:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=780s) goes oh that's still broken it goes back until it passes so it fixes it. Now just [13:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=783s) until it passes so it fixes it. Now just [13:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=784s) until it passes so it fixes it. Now just one thing um when a test fails by [13:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=786s) one thing um when a test fails by [13:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=786s) one thing um when a test fails by default most tools they exit with a one [13:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=789s) default most tools they exit with a one [13:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=789s) default most tools they exit with a one exit code um but with a one that task by [13:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=792s) exit code um but with a one that task by [13:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=792s) exit code um but with a one that task by clock code just gets marked as completed [13:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=794s) clock code just gets marked as completed [13:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=794s) clock code just gets marked as completed anyway. So, make sure that your scripts [13:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=797s) anyway. So, make sure that your scripts [13:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=797s) anyway. So, make sure that your scripts exit with a twoexit code because that's [13:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=799s) exit with a twoexit code because that's [13:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=799s) exit with a twoexit code because that's what actually sends Cloud back to fix [13:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=801s) what actually sends Cloud back to fix [13:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=801s) what actually sends Cloud back to fix it. So, okay, now we have one session [13:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=805s) it. So, okay, now we have one session [13:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=805s) it. So, okay, now we have one session that I'm pretty comfortable with leaving [13:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=807s) that I'm pretty comfortable with leaving [13:27](https://www.youtube.com/watch?v=vlfCjWuv4es&t=807s) that I'm pretty comfortable with leaving alone. So, now we can start with a few [13:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=810s) alone. So, now we can start with a few [13:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=810s) alone. So, now we can start with a few more of those because they've all got [13:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=811s) more of those because they've all got [13:31](https://www.youtube.com/watch?v=vlfCjWuv4es&t=811s) more of those because they've all got their own machine. So, they're, you [13:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=813s) their own machine. So, they're, you [13:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=813s) their own machine. So, they're, you know, not all working in the same folder [13:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=815s) know, not all working in the same folder [13:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=815s) know, not all working in the same folder and we don't really have to think about [13:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=816s) and we don't really have to think about [13:36](https://www.youtube.com/watch?v=vlfCjWuv4es&t=816s) and we don't really have to think about work trees too much ourselves anymore. [13:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=819s) work trees too much ourselves anymore. [13:39](https://www.youtube.com/watch?v=vlfCjWuv4es&t=819s) work trees too much ourselves anymore. But then we still start every session [13:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=821s) But then we still start every session [13:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=821s) But then we still start every session ourselves and we review every one of [13:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=823s) ourselves and we review every one of [13:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=823s) ourselves and we review every one of them. And honestly, this is already a [13:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=825s) them. And honestly, this is already a [13:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=825s) them. And honestly, this is already a pretty good place to be. We are still [13:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=827s) pretty good place to be. We are still [13:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=827s) pretty good place to be. We are still fully in control over what gets started [13:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=830s) fully in control over what gets started [13:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=830s) fully in control over what gets started and when we call a task done. But even [13:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=834s) and when we call a task done. But even [13:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=834s) and when we call a task done. But even there, um, you know, we're still the [13:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=835s) there, um, you know, we're still the [13:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=835s) there, um, you know, we're still the ones kicking off every session and we're [13:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=837s) ones kicking off every session and we're [13:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=838s) ones kicking off every session and we're still in the loop for all of them. So, I [14:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=840s) still in the loop for all of them. So, I [14:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=840s) still in the loop for all of them. So, I think the most important thing to have [14:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=841s) think the most important thing to have [14:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=841s) think the most important thing to have in place here now is your CI because all [14:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=844s) in place here now is your CI because all [14:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=844s) in place here now is your CI because all of these sessions eventually open a PR [14:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=846s) of these sessions eventually open a PR [14:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=846s) of these sessions eventually open a PR and your CI runs the same checks on [14:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=848s) and your CI runs the same checks on [14:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=848s) and your CI runs the same checks on every one of them, no matter how that [14:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=850s) every one of them, no matter how that [14:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=850s) every one of them, no matter how that session was set up or, you know, what [14:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=851s) session was set up or, you know, what [14:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=851s) session was set up or, you know, what Claude decided to do along the way. And [14:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=853s) Claude decided to do along the way. And [14:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=853s) Claude decided to do along the way. And if you turn on autofix, we have like CI [14:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=855s) if you turn on autofix, we have like CI [14:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=855s) if you turn on autofix, we have like CI autofix. Claude can also see when one of [14:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=858s) autofix. Claude can also see when one of [14:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=858s) autofix. Claude can also see when one of those checks fails and then go back to [14:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=860s) those checks fails and then go back to [14:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=860s) those checks fails and then go back to fix it. So you're not the one spending [14:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=862s) fix it. So you're not the one spending [14:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=862s) fix it. So you're not the one spending your review on things that you know a [14:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=864s) your review on things that you know a [14:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=864s) your review on things that you know a check could have already caught. [14:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=866s) check could have already caught. [14:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=866s) check could have already caught. And for some of them, you can even turn [14:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=868s) And for some of them, you can even turn [14:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=868s) And for some of them, you can even turn on merge when ready. So then GitHub [14:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=870s) on merge when ready. So then GitHub [14:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=870s) on merge when ready. So then GitHub merges a PR once all of the checks have [14:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=873s) merges a PR once all of the checks have [14:33](https://www.youtube.com/watch?v=vlfCjWuv4es&t=873s) merges a PR once all of the checks have passed. So you're not in a loop to [14:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=874s) passed. So you're not in a loop to [14:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=874s) passed. So you're not in a loop to review. Um but again, that that's not [14:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=877s) review. Um but again, that that's not [14:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=877s) review. Um but again, that that's not the default and that really depends on [14:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=878s) the default and that really depends on [14:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=878s) the default and that really depends on how risky the change is. So for example [14:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=880s) how risky the change is. So for example [14:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=880s) how risky the change is. So for example on our team uh every PR gets reviewed by [14:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=883s) on our team uh every PR gets reviewed by [14:43](https://www.youtube.com/watch?v=vlfCjWuv4es&t=883s) on our team uh every PR gets reviewed by cloud first. So not by the session that [14:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=885s) cloud first. So not by the session that [14:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=885s) cloud first. So not by the session that wrote it. So it comes with a fresh [14:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=888s) wrote it. So it comes with a fresh [14:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=888s) wrote it. So it comes with a fresh context and also a separate group of [14:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=890s) context and also a separate group of [14:50](https://www.youtube.com/watch?v=vlfCjWuv4es&t=890s) context and also a separate group of agents to just double check every bug it [14:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=892s) agents to just double check every bug it [14:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=892s) agents to just double check every bug it finds. And then how much we review [14:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=895s) finds. And then how much we review [14:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=895s) finds. And then how much we review ourselves just depends on the risk. So [14:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=897s) ourselves just depends on the risk. So [14:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=897s) ourselves just depends on the risk. So smaller changes like docer copy, they [14:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=899s) smaller changes like docer copy, they [14:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=899s) smaller changes like docer copy, they don't need a person. So once the [15:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=900s) don't need a person. So once the [15:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=900s) don't need a person. So once the automated checks are green, we're pretty [15:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=902s) automated checks are green, we're pretty [15:02](https://www.youtube.com/watch?v=vlfCjWuv4es&t=902s) automated checks are green, we're pretty good to go. um maybe ordinary product [15:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=905s) good to go. um maybe ordinary product [15:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=905s) good to go. um maybe ordinary product changes just need one reviewer and then [15:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=907s) changes just need one reviewer and then [15:07](https://www.youtube.com/watch?v=vlfCjWuv4es&t=907s) changes just need one reviewer and then risky changes might need two. Now [15:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=910s) risky changes might need two. Now [15:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=910s) risky changes might need two. Now ultimately kind of where we see things [15:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=912s) ultimately kind of where we see things [15:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=912s) ultimately kind of where we see things going um and we're already kind of doing [15:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=915s) going um and we're already kind of doing [15:15](https://www.youtube.com/watch?v=vlfCjWuv4es&t=915s) going um and we're already kind of doing more of that at anthropic is that you're [15:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=916s) more of that at anthropic is that you're [15:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=916s) more of that at anthropic is that you're not really starting all of those [15:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=918s) not really starting all of those [15:18](https://www.youtube.com/watch?v=vlfCjWuv4es&t=918s) not really starting all of those sessions yourself either because if [15:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=919s) sessions yourself either because if [15:20](https://www.youtube.com/watch?v=vlfCjWuv4es&t=920s) sessions yourself either because if you're anything like me, I'm terrible at [15:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=921s) you're anything like me, I'm terrible at [15:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=921s) you're anything like me, I'm terrible at context switching. So having to manage [15:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=923s) context switching. So having to manage [15:23](https://www.youtube.com/watch?v=vlfCjWuv4es&t=923s) context switching. So having to manage so many sessions at the at the same time [15:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=925s) so many sessions at the at the same time [15:25](https://www.youtube.com/watch?v=vlfCjWuv4es&t=925s) so many sessions at the at the same time doesn't always go well. So instead you [15:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=928s) doesn't always go well. So instead you [15:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=928s) doesn't always go well. So instead you just have one main session that's your [15:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=930s) just have one main session that's your [15:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=930s) just have one main session that's your orchestrator. And you can already see [15:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=932s) orchestrator. And you can already see [15:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=932s) orchestrator. And you can already see this in things like cloud code projects [15:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=934s) this in things like cloud code projects [15:34](https://www.youtube.com/watch?v=vlfCjWuv4es&t=934s) this in things like cloud code projects which is still in beta but you just tell [15:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=937s) which is still in beta but you just tell [15:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=938s) which is still in beta but you just tell it what it needs to be done and the [15:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=940s) it what it needs to be done and the [15:40](https://www.youtube.com/watch?v=vlfCjWuv4es&t=940s) it what it needs to be done and the orchestrator splits that up and then [15:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=942s) orchestrator splits that up and then [15:42](https://www.youtube.com/watch?v=vlfCjWuv4es&t=942s) orchestrator splits that up and then starts other sessions for each task and [15:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=944s) starts other sessions for each task and [15:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=944s) starts other sessions for each task and it keeps track of what comes back. So it [15:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=946s) it keeps track of what comes back. So it [15:46](https://www.youtube.com/watch?v=vlfCjWuv4es&t=946s) it keeps track of what comes back. So it automatically starts and verifies these [15:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=948s) automatically starts and verifies these [15:48](https://www.youtube.com/watch?v=vlfCjWuv4es&t=948s) automatically starts and verifies these threads which are just new sessions. So [15:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=951s) threads which are just new sessions. So [15:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=951s) threads which are just new sessions. So instead of talking to eight sessions, [15:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=952s) instead of talking to eight sessions, [15:52](https://www.youtube.com/watch?v=vlfCjWuv4es&t=952s) instead of talking to eight sessions, you're just talking to one. And the main [15:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=955s) you're just talking to one. And the main [15:55](https://www.youtube.com/watch?v=vlfCjWuv4es&t=955s) you're just talking to one. And the main session doesn't write the code itself. [15:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=957s) session doesn't write the code itself. [15:57](https://www.youtube.com/watch?v=vlfCjWuv4es&t=957s) session doesn't write the code itself. So there you want a model that's really [15:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=959s) So there you want a model that's really [15:59](https://www.youtube.com/watch?v=vlfCjWuv4es&t=959s) So there you want a model that's really good at deciding what to do and judging [16:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=961s) good at deciding what to do and judging [16:01](https://www.youtube.com/watch?v=vlfCjWuv4es&t=961s) good at deciding what to do and judging what comes back. A model like Fable and [16:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=964s) what comes back. A model like Fable and [16:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=964s) what comes back. A model like Fable and if something comes back that isn't [16:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=965s) if something comes back that isn't [16:05](https://www.youtube.com/watch?v=vlfCjWuv4es&t=965s) if something comes back that isn't actually done, it just, you know, sends [16:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=966s) actually done, it just, you know, sends [16:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=966s) actually done, it just, you know, sends it back. So until it's actually done and [16:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=969s) it back. So until it's actually done and [16:09](https://www.youtube.com/watch?v=vlfCjWuv4es&t=969s) it back. So until it's actually done and passes everything and the threads need a [16:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=971s) passes everything and the threads need a [16:11](https://www.youtube.com/watch?v=vlfCjWuv4es&t=971s) passes everything and the threads need a model that's good at actually doing the [16:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=973s) model that's good at actually doing the [16:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=973s) model that's good at actually doing the work like Opus 55. And you can set those [16:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=976s) work like Opus 55. And you can set those [16:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=976s) work like Opus 55. And you can set those separately. So you know, one model for [16:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=977s) separately. So you know, one model for [16:17](https://www.youtube.com/watch?v=vlfCjWuv4es&t=977s) separately. So you know, one model for the main session and then another one [16:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=979s) the main session and then another one [16:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=979s) the main session and then another one for the threads. And another benefit is [16:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=981s) for the threads. And another benefit is [16:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=982s) for the threads. And another benefit is that with this setup with code projects [16:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=984s) that with this setup with code projects [16:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=984s) that with this setup with code projects specifically, we don't manage context [16:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=986s) specifically, we don't manage context [16:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=986s) specifically, we don't manage context anymore. That's all done automatically [16:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=988s) anymore. That's all done automatically [16:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=988s) anymore. That's all done automatically for us. So, we no longer have to think [16:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=990s) for us. So, we no longer have to think [16:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=990s) for us. So, we no longer have to think about, you know, when to compact or [16:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=992s) about, you know, when to compact or [16:32](https://www.youtube.com/watch?v=vlfCjWuv4es&t=992s) about, you know, when to compact or clear all of that stuff. Okay. So, if [16:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=995s) clear all of that stuff. Okay. So, if [16:35](https://www.youtube.com/watch?v=vlfCjWuv4es&t=995s) clear all of that stuff. Okay. So, if you look back at those old best [16:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=997s) you look back at those old best [16:37](https://www.youtube.com/watch?v=vlfCjWuv4es&t=997s) you look back at those old best practices, they're all still there, but [16:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=998s) practices, they're all still there, but [16:38](https://www.youtube.com/watch?v=vlfCjWuv4es&t=998s) practices, they're all still there, but they've just moved. So, the model steers [16:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1001s) they've just moved. So, the model steers [16:41](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1001s) they've just moved. So, the model steers it now. Um, auto mode approves your deny [16:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1004s) it now. Um, auto mode approves your deny [16:44](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1004s) it now. Um, auto mode approves your deny rules and the environment restrict [16:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1005s) rules and the environment restrict [16:45](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1005s) rules and the environment restrict what's possible. Um, your test, your [16:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1007s) what's possible. Um, your test, your [16:47](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1007s) what's possible. Um, your test, your hooks, and your CI checks it. And then [16:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1009s) hooks, and your CI checks it. And then [16:49](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1009s) hooks, and your CI checks it. And then ultimately, cloud can just start them as [16:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1011s) ultimately, cloud can just start them as [16:51](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1011s) ultimately, cloud can just start them as well. So almost everything we used to do [16:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1014s) well. So almost everything we used to do [16:54](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1014s) well. So almost everything we used to do ourselves is now either the model, the [16:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1016s) ourselves is now either the model, the [16:56](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1016s) ourselves is now either the model, the harness or the environment. And we're [16:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1018s) harness or the environment. And we're [16:58](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1018s) harness or the environment. And we're really just here to start that first one [17:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1020s) really just here to start that first one [17:00](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1020s) really just here to start that first one and to just review what comes out. So if [17:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1023s) and to just review what comes out. So if [17:03](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1023s) and to just review what comes out. So if you want to try this in your own setup, [17:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1024s) you want to try this in your own setup, [17:04](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1024s) you want to try this in your own setup, I would say just take one thing that's [17:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1026s) I would say just take one thing that's [17:06](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1026s) I would say just take one thing that's way bigger than you normally give [17:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1028s) way bigger than you normally give [17:08](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1028s) way bigger than you normally give Claude. Make sure your CI actually [17:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1030s) Claude. Make sure your CI actually [17:10](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1030s) Claude. Make sure your CI actually catches the things you care about and [17:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1032s) catches the things you care about and [17:12](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1032s) catches the things you care about and then spend a bit of time just [17:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1033s) then spend a bit of time just [17:13](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1033s) then spend a bit of time just configuring your environment like its [17:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1034s) configuring your environment like its [17:14](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1034s) configuring your environment like its network access, its secret, and its, you [17:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1036s) network access, its secret, and its, you [17:16](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1036s) network access, its secret, and its, you know, setup scripts. and then start in [17:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1039s) know, setup scripts. and then start in [17:19](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1039s) know, setup scripts. and then start in the cloud and see where your sessions [17:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1041s) the cloud and see where your sessions [17:21](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1041s) the cloud and see where your sessions get stuck and where you're still in the [17:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1042s) get stuck and where you're still in the [17:22](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1042s) get stuck and where you're still in the loop because that's the next thing to [17:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1044s) loop because that's the next thing to [17:24](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1044s) loop because that's the next thing to hand over to Claude. I also highly [17:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1046s) hand over to Claude. I also highly [17:26](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1046s) hand over to Claude. I also highly recommend checking out our best [17:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1048s) recommend checking out our best [17:28](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1048s) recommend checking out our best practices in the docs as well. Thank you [17:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1050s) practices in the docs as well. Thank you [17:30](https://www.youtube.com/watch?v=vlfCjWuv4es&t=1050s) practices in the docs as well. Thank you so much.