# Cognitive Exhaust Fumes, or: Read-Only AI Is Underrated — Šimon Podhajský, Head of AI, Waypoint

Šimon Podhajský

AI Engineer Europe 2026

[Talk page](https://ai.engineer/talks/u0TOSBbAw7c-cognitive-exhaust-fumes-read-only-ai-is) · [Video](https://www.youtube.com/watch?v=u0TOSBbAw7c)

## Cognitive Exhaust Fumes, or: Read-Only AI Is Underrated

A personal AI can connect your reading, commitments, and relationships without acting for you. The useful boundary separates observation from action, but it does not eliminate privacy risk.

### What can an AI learn without acting for you?

[0:01](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=1s)

What would a personal AI be useful for if it knew you but could not send a message, change a task, or act on your behalf? Šimon Podhajský starts with that constraint: six data sources, read access only, and no permissions to write back. The restriction is intentional, not an unfinished step toward automation.

He calls the raw material **cognitive exhaust fumes**: digital activity left behind as a byproduct of thinking. A single trace may look disposable. Taken together, those traces can reveal patterns in the activity that produced them, much as analyzing exhaust can help diagnose an engine.

The useful patterns cross application boundaries:

- **Intention-action gaps:** differences between what you intend to do and what your activity shows.
- **Attention drift:** changes in what captures your attention.
- **Relationship decay:** connections that receive less attention over time.

Your email client does not know what you journaled; your task manager does not know what you browsed. **The cross-source signal is the product.**

![“What the Exhaust Reveals” slide with three example boxes and the data sources supporting each.](https://ai.engineer/api/talk-essay-media/42f2a56d022cb9b27e15aca870839f1fe908739c933a5ab4ed1b859715c8e073)

Cross-source signals reveal intention-action gaps, attention drift, and relationship decay.

### Separate sources, analysis, and outputs

[1:29](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=89s)

Read-only does not mean the system cannot create a file. It means the AI cannot modify the sources it observes. Podhajský separates the system into three zones:

| Zone | Purpose | Write boundary |
|---|---|---|
| Sources | Supply personal activity | No AI writes back |
| Workspace | Analyze the collected material | Analysis happens here |
| Outputs | Hold results for human review | Separate from the sources |

His outputs land in a separate Obsidian vault. A separate Notion destination or ordinary text files would serve the same purpose: analysis becomes something to review without rewriting the underlying record.

### Turn a week of activity into reflection

[1:54](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=114s)

The first application is a weekly reflection inspired by David Allen’s [Getting Things Done Weekly Review](https://gettingthingsdone.com/2018/08/episode-43-the-power-of-the-gtd-weekly-review/). Instead of asking the user to reconstruct the week from memory, the system synthesizes material from all six sources. The result can be an occasionally brutal account of how the week was spent.

The workflow runs through **Claude**, with the logic packaged as a slash command or [skill](https://code.claude.com/docs/en/skills):

1. Invoke the weekly reflection skill.
2. Launch a Python script that gathers data from the read-only sources.
3. Apply prepared prompts and call the Anthropic API for structured outputs.
4. Turn those outputs into a Markdown document for review.

The separation between structured analysis and document rendering makes the final artifact straightforward to inspect. A small Python renderer illustrates that last step, taking already-produced analysis and writing only to a separate output directory:

```python
from pathlib import Path


def write_reflection(sections: dict[str, list[str]], output_dir: Path) -> Path:
    output_dir.mkdir(parents=True, exist_ok=True)
    document = ["# Weekly reflection"]
    for heading, items in sections.items():
        document.append(f"## {heading}")
        document.append("\n".join(f"- {item}" for item in items))

    destination = output_dir / "weekly-reflection.md"
    destination.write_text("\n\n".join(document) + "\n", encoding="utf-8")
    return destination
```

This output convention expresses the architecture; source permissions must enforce the read-only boundary separately.

When the run finishes, Podhajský opens the Markdown in Cursor and switches to a readable preview. It identifies themes of the week, tensions and conflicts, commitments and relationships, notable moments, and questions for reflection. The relationship and commitment sections are revealing partly through omissions: what did not receive attention can matter as much as what did. The aim is to reflect on how he is thinking, rather than produce a productivity score.

### Find someone to discuss your reading with

[3:38](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=218s)

The next question is more specific: given what I have been reading, who in my network should I discuss it with? Podhajský enjoys sharing reading with others but wants to look beyond the same three habitual recipients. This example combines four sources that were not designed to work together. Its output is a suggestion: nothing is sent and nothing is scheduled.

A plain-language request activates the cross-source query skill. It works through curated databases that receive regular ingestion, including a Vivaldi SQLite database used to examine reading activity. The query identifies articles he has read frequently and articles still open in tabs, then looks for people whose profiles suggest an interest in those topics.

Clay MCP supplies the relationship search. Podhajský describes Clay as his CRM—or, more personally, his friend relationship system. In this run, Claude searches for contacts interested in AI, European tech, and education. He calls this the weakest part of the workflow because the Clay searches are slow. The read-only constraint belongs to his system’s configuration; it should not be inferred merely from the connector’s name.

![Claude Code terminal showing article query output, Clay contact searches for AI, European tech, and education, and a returned contact response with a large-response warning.](https://ai.engineer/api/talk-essay-media/0c49458a0bac2faae5e1e6d2e7bae0228225038ca48e4ad3dcc6201c67d3cec2)

Claude searches Clay contacts by topic after retrieving recent reading activity.

### From large search responses to specific recommendations

[5:42](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=342s)

The search results consume substantial context. Podhajský recommends starting with a clean session and clearing it afterward; he refers to a one-million-token context window for “4.6.” That is a capacity reference, not measured token usage or retrieval accuracy. Both [**Claude Opus 4.6**](https://www.anthropic.com/news/claude-opus-4-6) and [**Claude Sonnet 4.6**](https://www.anthropic.com/news/claude-sonnet-4-6) were announced with that capacity, so the shorthand does not identify the demo’s exact variant or enabled limit.

Once the Clay responses arrive, Claude synthesizes the candidate contacts and maps each person to one article. As the demonstration continues, that mapping is still being assembled with shell commands in Claude Code. Podhajský mentions [auto mode](https://claude.com/blog/auto-mode) and `--dangerously-skip-permissions` as ways to reduce permission friction. They have different protections: auto mode makes permission decisions with safeguards, while the flag bypasses permission checks. Neither establishes the source-access boundary.

The finished recommendations look useful to him: people he might want to approach, whom he has not yet talked to about these readings. The result even finds an article’s author already in his network. In the displayed table, Jiří Dolejš appears with the rationale “He wrote it,” alongside contact, rationale, and opener columns. The proposed conversation remains for Podhajský to initiate.

![Claude Code results show contact, rationale, and opener columns; Jiří Dolejš’s rationale reads “He wrote it,” followed by a highest-value moves summary.](https://ai.engineer/api/talk-essay-media/396847db3d9e64fdb80eff6a4fd20a7d9674e4c101db8879117401bf7897ee46)

The recommendations include an article’s author as a conversation partner.

The browser alone could identify reading activity but would lack the contacts. The CRM could find people but would lack the reading context. Joining those records produces a person-to-article recommendation that neither source can supply on its own.

### Keep action—and the feedback loop—with the user

[7:28](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=448s)

Why stop at a recommendation when the system could send the message? Podhajský’s answer begins with asymmetric consequences. He calls the downside of an incorrect read-only suggestion zero when he can simply ignore it. An incorrect write can instead affect a relationship, a career, or a reputation. That comparison concerns the consequences of acting on someone’s behalf; it does not establish that reading private data is risk-free. He would rather forgo automated emails than accept a consequential misfire.

The second reason concerns the quality of the evidence. If AI starts writing into the same systems it analyzes, future observations mix human activity with AI-generated activity. A pattern in the record becomes harder to attribute: is this how the person thinks, or how the automation behaves? **Keeping sources untouched preserves a clearer record of the activity being observed.**

An observer still changes behavior. Reading a reflection may prompt a different decision next week. But the loop runs through the person: read, decide, then act. It does not automatically close through an AI rewrite. Podhajský extends this into a preference for reclaiming personal agency, even in drafting—while acknowledging that this preference may be a hard sell to an audience building AI tools.

### An observer serves a different need

[8:44](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=524s)

Podhajský has also tried [OpenClaw](https://docs.openclaw.ai/gateway/sandboxing) with a read-only mount. His preference for a dedicated observer comes from the value he experiences per interaction. His illustrative comparison is an agent saving thirty seconds on a weather check versus an observer revealing two weeks of avoiding his most important project. This is his assessment of usefulness, not a measured comparison between products.

He also sees less opportunity for exfiltration and less contamination of the observed record, although the security boundary needs closer examination. The product distinction stands independently: agents perform actions; observers help people notice and interpret patterns. Read-only need not be a capability users eventually graduate from. As he puts it, “A mirror isn't a broken butler.”

### The useful combination is also a sensitive target

[9:40](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=580s)

The same cross-referencing that makes the observer valuable creates its central privacy risk. The **mosaic effect** is what happens when many small pieces of information combine into a revealing picture. A fragment that seems modest in isolation may become sensitive when connected to other fragments. Podhajský’s own product description therefore doubles as a threat description: the combined view is an attractive target precisely because it reveals more than any single source. Read-only access protects the source record from modification; it does not make that combined picture confidential.

![“The Mosaic Effect” slide contrasts four low-sensitivity sources with a high-sensitivity cross-referenced profile and states that read-only protects integrity, not confidentiality.](https://ai.engineer/api/talk-essay-media/60d20eb54e807bf3f0fb13c5c9c04d71fa6fa2ad73657c70d0e33239929644db)

The mosaic effect: individually modest signals combine into sensitive inferences.

### Read-only does not close the exfiltration path

[10:11](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=611s)

Simon Willison’s [lethal trifecta](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/) makes the remaining risk explicit: **private data**, **untrusted content**, and **external communication** coexist in one system. Podhajský initially thought read-only access broke that combination. He then recognized that removing source writes only removes some natural exfiltration channels. Shell access can still communicate externally, so the third capability remains.

Even in ordinary operation, this system sends data to Anthropic. Podhajský describes a mostly open network and more accessible information than the task strictly requires. He does not claim the system is secure or fireproof. His position is that he has examined its exposure and decided which risks he is willing to carry—a different posture from leaving them unrecognized.

That leaves a useful, bounded purpose for the system: turn an underused personal dataset into material for reflection. Your digital activity already records traces of attention, commitments, and relationships. Bringing those traces together can help you decide what to change, while leaving the decision and the action with you.

## Resources

- [Simon Willison's lethal trifecta](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/) — Explains how private data, untrusted inputs, and external communication combine to enable data theft through prompt injection.
- [Claude Code skills documentation](https://code.claude.com/docs/en/skills) — Current documentation for extending Claude Code with reusable skills.
- [OpenClaw sandboxing and workspace access](https://docs.openclaw.ai/gateway/sandboxing) — Current sandbox configuration guidance, including read-only workspace access and separate mount permissions.
- [Claude Code auto mode](https://claude.com/blog/auto-mode) — Explains automated permission decisions and how they differ from disabling permission checks.
- [David Allen on the GTD Weekly Review](https://gettingthingsdone.com/2018/08/episode-43-the-power-of-the-gtd-weekly-review/) — David Allen's discussion of the Weekly Review, with eleven review steps and a checklist.
- [Claude Opus 4.6 launch announcement](https://www.anthropic.com/news/claude-opus-4-6) — The February 2026 announcement introducing Opus 4.6 and its one-million-token context window in beta.

## Slides

### Cognitive Exhaust Fumes

Šimon Podhajský introduces a personal AI that knows you without acting on your behalf, and discusses how read-only access can mitigate personal AI risks.

[0:00](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=0s) [Title slide: Cognitive Exhaust Fumes, What Read-Only AI Sees That You Can’t.](https://assets.ai.engineer/sites/aiecode2025/sha256/2832d9ee84d3ffaa795b82a824d0e3ca5c15df30be4eb11528b8e40edadda247.webp)

Cognitive Exhaust Fumes
What Read-Only AI Sees That You Can’t
Šimon Podhajský · ai.engineer/europe, April 2026

### Every personal AI demo looks a bit like this

Podhajský contrasts demos that send emails, schedule meetings, organize files and reply on Slack with his own system, which has no write permissions by design.

[0:27](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=27s) [Examples of AI acting on a user’s behalf, followed by the statement that this system cannot do those things.](https://assets.ai.engineer/sites/aiecode2025/sha256/b413237d2bbcdc037efab8b01d83ab257190b007408386f8533b4e82022ef675.webp)

Every personal AI demo looks a bit like this

"AI sent the email for me"
"AI scheduled my meeting"
"AI organized my files"
"AI replied to my Slack"

Mine can't do any of that.

### What I Built Instead

The system reads six sources, from three years of email to live browser tabs and a CRM. Its lack of write permissions is intentional.

[0:32](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=32s) [A read-only system connects email, journal, browser sessions, tasks, notes and contacts.](https://assets.ai.engineer/sites/aiecode2025/sha256/cdda7d4758a5c8c835f3ac2b8ff83b388a170059c078508584f2f486bb5f6b62.webp)

What I Built Instead

A system that connects to six data sources and can't write to any of them.

Email
3 years of archive

Journal
Daily entries

Browser Sessions
Live tab state

Tasks
Microsoft To-Do

Notes
Obsidian Vault

Contacts
CRM via MCP

### What are cognitive exhaust fumes?

Email, tabs, tasks and journal entries leave traces of thinking. Podhajský compares analyzing them together to diagnosing an engine from its exhaust.

[0:38](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=38s) [Digital activity is described as exhaust from thinking that can reveal patterns when analyzed together.](https://assets.ai.engineer/sites/aiecode2025/sha256/13b789a11ecdd8e80514a71527ffb67927a747f4ef2ccd3a9507ccedc4b00ea6.webp)

What are cognitive exhaust fumes?

Every email you send, tab you open, task you create and abandon, journal entry you write — that's the exhaust from your thinking.

Individually, each source is noise.
Analyzed together, they reveal the engine.

### What the Exhaust Reveals

Podhajský introduces examples of what cognitive exhaust can reveal, including intention-action gaps and attention drift.

[0:57](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=57s) [Title slide reading “What the Exhaust Reveals.”](https://assets.ai.engineer/sites/aiecode2025/sha256/315f324e5f6fab8d50507a2475a02e8d9fd6875dcd8e8c52e3650b90725f4dad.webp)

What the Exhaust Reveals

### What the Exhaust Reveals

Combining journals, tasks, browsing, email and CRM can reveal intentions without action, attention shifting away from a deadline, and neglected follow-ups—patterns no single source captures.

[1:08](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=68s) [Examples of intention-action gaps, attention drift, and relationship decay, each combining information from multiple sources.](https://assets.ai.engineer/sites/aiecode2025/sha256/b58d135a69002f01be543b3f5001d9961aee31a4dbfb7321e9f35464d21ca637.webp)

What the Exhaust Reveals

Intention-action gaps — "You journaled about writing that blog post three times. You haven't opened the draft once."
journal + tasks + notes

Attention drift — "Your browser tabs shifted from project A to project B mid-week. Project A's deadline is Friday."
browser sessions + tasks

Relationship decay — "You haven't emailed this collaborator in 6 weeks. Last time, you said you'd follow up."
email + CRM + journal

### The System

The talk turns from cross-source signals to a closer look at the system.

[1:24](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=84s) [The System title over an orange background with gears and circuitry.](https://assets.ai.engineer/sites/aiecode2025/sha256/6c7e930feae018d9dcf717892308ba1b58944de8d1c4c946311912d7bcff7d18.webp)

The System

### Three Zones

Claude Code reads six sources without writing back, analyzes them in a workspace, and writes weekly reflections, draft rankings, and alerts and briefings to an Obsidian vault for review.

[1:28](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=88s) [Diagram showing read-only sources feeding a Claude Code analysis workspace, which writes reflections, rankings, and briefings for review in Obsidian.](https://assets.ai.engineer/sites/aiecode2025/sha256/fc5ef4521acfdf37d8582b56bbff3a903cdd60b2120b34d7e8a83f65c63fd5f8.webp)

Three Zones

READ-ONLY SOURCES
Email (Maildir)
Journal (SQLite)
Browser (Vivaldi)
Tasks (Graph API)
Notes (Obsidian)
Contacts (Clay MCP)

READ

ANALYSIS WORKSPACE
Claude Code
18 Specialized Skills
Scheduled Analysis

WRITE

OUTPUTS
Weekly Reflections
Draft Rankings
Alerts & Briefings

Me
review in Obsidian

### Weekly Reflection

A weekly reflection combines six sources to compare stated priorities with behavior. The example contrasts 2h on apartment furnishing with 11h on side projects, alongside relationship, task and browser activity.

[2:02](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=122s) [Slide describes a weekly reflection and gives an example of synthesizing activity across sources.](https://assets.ai.engineer/sites/aiecode2025/sha256/589303e828cfb2afd44056387692f67734ebc49f01b9b0d3c48bf1df68fa568c.webp)

In Practice: Weekly Reflection

"A mirror for how you spent your week."

Six sources, one synthesis. No single source gives you this picture.

You said apartment furnishing matters but spent 2h on it vs. 11h on side projects. Three people you called important haven't heard from you in a month. You created 12 tasks and completed 4. Your browser has 3 abandoned research threads from Monday.

### Starting the Reflection

The speaker starts the weekly reflection in Claude with /weekly-reflection. He says the skill launches a Python script to process read-only sources and produce structured outputs for a Markdown review.

[2:12](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=132s) [Terminal shows the weekly reflection command composing.](https://assets.ai.engineer/sites/aiecode2025/sha256/d129b8f9d4677518f7fd679d2bf4a7a8c3561cb63d8707d633388478d663891e.webp)

Claude Code v2.1.81
Opus 4.6 (1M context) · Claude Max
~/Documents/Claude-Vault

> /weekly-reflection

* Composing...

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24.3. 9:02
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
~/D/Claude-Vault on main [illegible] is v0.1.0 via v3.13.3 (Claude-Vault)
> claude
Claude-Vault | Opus 4.6 (1M context) | 0% | (main | 1 files +1 -1 ) [$0.00]  medium · /effort
⌘K to generate command

### Running the Reflection

The weekly reflection skill runs a Python script that gathers read-only data and sends prepared prompts to the Anthropic API for a Markdown reflection.

[2:17](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=137s) [Terminal shows the reflection skill running a Python script.](https://assets.ai.engineer/sites/aiecode2025/sha256/be893f2a33a97fb946b3cb5b693b9754d07d0bf1bfd6a3d5bdd432e3308700f1.webp)

~/D/Claude-Vault on main [!↑] is v0.1.0 via v3.13.3 (Claude-Vault)
❯ claude

Claude Code v2.1.81
Opus 4.6 (1M context) · Claude Max
~/Documents/Claude-Vault

❯ /weekly-reflection

Running the weekly reflection skill.

Bash(cd ~/Documents/Claude-Vault/skills/weekly-reflection && uv run python3 weekly_reflection.py --redact 2>&1)
  warning: `VIRTUAL_ENV=/Users/simon/Documents/Claude-Vault/.venv` does not match the project environment path `.venv` and will be
  (45s · timeout 5m)
  (ctrl+b to run in background)

✻ Booping… (51s · ↓ 60 tokens)

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24. 3. 9:03
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
Claude-Vault | Opus 4.6 (1M context) | 2% | (main | 1 files +1 -1 ) [$0.14]
⌘K to generate command
Claude-Vault  main*  Claude-Vault  0  0  Cursor Tab

### Weekly Reflection

The completed reflection identifies four themes, then prompts Šimon to examine priorities, relationships, and unfinished decisions in a Markdown preview.

[2:55](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=175s) [Terminal reports a saved weekly reflection and lists four themes.](https://assets.ai.engineer/sites/aiecode2025/sha256/12f43c90b43b5f71b347d5c797215ccd13fcc6c290c371114891218ff5262ec1.webp)

> claude
Claude Code v2.1.81
Opus 4.6 (1M context) · Claude Max
~/Documents/Claude-Vault

> /weekly-reflection

Running the weekly reflection skill.

Bash(cd ~/Documents/Claude-Vault/skills/weekly-reflection && uv run python3 weekly_reflection.py --redact 2>&1)
  warning: `VIRTUAL_ENV=/Users/simon/Documents/Claude-Vault/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
  ✓ Found statistical review: 2026-03-17-weekly-review.md
  … +43 lines (ctrl+o to expand)
  (timeout 5m)

Read 1 file (ctrl+o to expand)

Weekly reflection generated and saved to output/reflections/2026-03-24-weekly-reflection.md.

Highlights:
- 4 themes: High-velocity task execution (14/14), AI educator identity building, Waypoint technical deepening, post-retreat relationship activation

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24.3. 9:03
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
~/D/Claude-Vault on main [!↑] is v0.1.0 via v3.13.3 (Claude-Vault)
weekly-reflection | Opus 4.6 (1M context) | 3% | (main | 1 files +1 -1 ) [$0.16]
⌘K to generate command
Claude-Vault  main*  Claude-Vault  0  0  Cursor Tab

[3:25](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=205s) [A Markdown preview shows reflection questions about priorities, follow-ups, and a draft.](https://assets.ai.engineer/sites/aiecode2025/sha256/6accc9ffdc6ffb9c8cb916c6b2aa05695342ca4f5dafe027823768fcd2262763.webp)

technical knowledge — it might be that none of the options feel clean, or you're waiting for input from someone else.

2. Of the many roles you played this week (engineer, educator, community builder, content creator, financial planner), which one would you protect if you could only keep two?
Your task list spans at least six distinct domains. The throughput was great this week, but bandwidth is finite. Knowing your priority hierarchy before a crunch week hits is more useful than figuring it out during one.

3. What came out of the financial strategy session with Claude that you want to act on?
You completed the task, but deep research is only valuable if it leads to decisions or behavior changes. Was there a concrete next step, or does this need a follow-up session?

4. Are the víkendovka follow-ups seeds you want to water, or was the follow-up itself the full gesture?
You completed the follow-ups quickly, which is good social hygiene. But some of these connections might deserve deeper investment — especially if they connect to your evals.cz or AI community work.

5. The LinkedIn draft about AI experts — are you ready to publish it, or is it sitting in drafts because the framing isn't quite right yet?
Both LinkedIn drafts are in 'draft' status. The AI expert piece feels like it has real conviction behind it, but drafts that linger too long often lose their energy. The Czech AI discourse moves fast.

This reflection is generated from the statistical review and actual text from your week.

For pure stats, see [[reviews/2026-03-17-weekly-review]]

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24.3. 9:04
Claude-Vault
2026-03-24-weekly-reflection.md
output > reflections > 2026-03-24-weekly-reflection.md
Preview  Markdown
Options  Capture
Claude-Vault  main*  Claude-Vault  0  0  Cursor Tab

### In Practice: Cross-Source Query

A cross-source query connects recent reading and notes with CRM and email context to suggest people to talk to. The example flags two relationships going cold; nothing is sent or scheduled.

[3:52](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=232s) [A slide combines browser sessions, notes, CRM, and email to suggest relevant people to contact about recent reading.](https://assets.ai.engineer/sites/aiecode2025/sha256/2e989283187665442bc30cf6f495f90bc124b9316da377861c866afc6400e2e9.webp)

In Practice: Cross-Source Query

Sources crossed:
• Browser sessions — what you’ve been reading
• Notes — what you’ve been thinking about
• CRM — who you know, when you last spoke
• Email — relationship recency and additional context

You've been reading about eval frameworks all week. Three people in your network work on this: A (last email: 2 months ago), B (met at a conference, never followed up), C (active collaborator). A and B are going cold.

### Cross-source query

A plain-language request activates a cross-source skill, which plans to pull recent reading history and contacts in parallel to find people to discuss the articles with.

[4:28](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=268s) [A terminal shows a question about matching recent articles with contacts and the cross-source skill starting.](https://assets.ai.engineer/sites/aiecode2025/sha256/4e014134ed4d356bd7f0b1918bc887ad08c808062c3374cf280ccff853556dea.webp)

Claude Code v2.1.81
Opus 4.6 (1M context) · Claude Max
~/Documents/Claude-Vault

I want to talk to some people about the articles I've read this past week. Who in my network should I talk to about which articles?

Skill(cross-source-queries)
Successfully loaded skill

Let me pull your recent reading history and contacts in parallel.

Synthesizing…

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24. 3. 9:07
Work  Off
Terminal  Problems  Output  Debug Console  Ports
~/D/Claude-Vault on main [!] is v0.1.0 via v3.13.3 (Claude-Vault)
claude
claude
claude
Claude-Vault | Opus 4.6 (1M context) | 2% | (main | 1 files +1 -1 ) [$0.09]
⌘K to generate command
Claude-Vault  main*  Claude-Vault  0  0  Cursor Tab  Ln 1, Col 1  Spaces: 3  UTF-8  LF  Markdown  Prettier

### Querying recent articles

Claude finds the initial Vivaldi history database empty, inspects another database, and queries its articles seen since March 17. The demo uses this reading history to find relevant contacts.

[5:01](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=301s) [The terminal shows database inspection followed by a query for recent articles.](https://assets.ai.engineer/sites/aiecode2025/sha256/b2f65bafb7da3b02ce8e86d7b71a39970b161a40d7b6731f9e68fb2f2d2bcb74.webp)

Skill(cross-source-queries)
  Successfully loaded skill

Let me pull your recent reading history and contacts in parallel.

Bash(sqlite3 ~/data/vivaldi/vivaldi_history.db "SELECT url, title, last_visit_time FROM urls WHERE last_visit_time
  > strftime('%s', '2026-03-17') * 1000000 + 1164447…)
  FALLBACK

Bash(ls ~/data/vivaldi/ 2>/dev/null && file ~/data/vivaldi/*.db 2>/dev/null; sqlite3
  ~/data/vivaldi/vivaldi_history.db ".tables" 2>/dev/null; sqlite3 ~/data/vivaldi/…)
  vivaldi.db
  vivaldi_history.db
  /Users/simon/data/vivaldi/vivaldi_history.db: empty
  … +2 lines (ctrl+o to expand)

Bash(sqlite3 ~/data/vivaldi/vivaldi.db ".tables" && sqlite3 ~/data/vivaldi/vivaldi.db ".schema" | head -80)
  articles        ingest_metadata
  CREATE TABLE articles (
                  id INTEGER PRIMARY KEY AUTOINCREMENT,
  … +24 lines (ctrl+o to expand)

Bash(sqlite3 ~/data/vivaldi/vivaldi.db "SELECT title, url, domain, last_seen, visit_count FROM articles WHERE
  last_seen >= '2026-03-17' ORDER BY last_seen DESC;")
  What is agentic engineering? – Agentic Engineering Patterns – Simon Willison's Weblog|https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/|simonwillison.net|2026-03-24|1
  Using incognito chats | Claude Help Center|https://support.claude.com/en/articles/12260368-using-incognito-
  … +110 lines (ctrl+o to expand)

Orbiting… (58s · ↓ 291 tokens)

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24. 3. 9:07
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
claude
Claude-Vault | Opus 4.6 (1M context) | 2% | (main | 1 files +1 -1 ) [$0.17]
⌘K to generate command
Claude-Vault  main*  Claude-Vault  Cursor Tab  Ln 1, Col 1  Spaces: 3  UTF-8  LF  Markdown  Prettier

### Matching recent articles with contacts

Claude searches Clay for contacts interested in AI, European tech, and education, then requests approval to extract contact details before mapping people to articles.

[5:31](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=331s) [The terminal shows contact searches for AI, European technology, and education topics.](https://assets.ai.engineer/sites/aiecode2025/sha256/be371f26f6208d778c275fcc938e36aa185b784b8f6e726647dd4a6a799e86ad.webp)

CREATE TABLE articles (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
… +24 lines (ctrl+o to expand)

Bash(sqlite3 ~/data/vivaldi/vivaldi.db "SELECT title, url, domain, last_seen, visit_count FROM articles WHERE last_seen >= '2026-03-17' ORDER BY last_seen DESC;")
  What is agentic engineering? - Agentic Engineering Patterns - Simon Willison's Weblog|https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/|simonwillison.net|2026-03-24|1
  Using incognito chats | Claude Help Center|https://support.claude.com/en/articles/12260368-using-incognito-
  … +110 lines (ctrl+o to expand)

Now let me search Clay for contacts relevant to these article topics.

clay - searchContacts (MCP)(query: "people interested in AI, LLMs, agentic engineering, or developer tools", job_title: ["AI Engineer","Machine Learning Engineer","Software Engineer","CTO","Head of AI","Data Scientist","Developer Advocate"], company_name: [], location: [], keywords: ["AI","LLM","agentic","evals","retrieval","NLP","developer tools"], limit: 20)
  Running…

clay - searchContacts (MCP)(query: "people in European tech, startups, or tech policy", job_title: ["Founder","CEO","Product Manager","Venture Capital","Tech Lead"], company_name: [], location: ["Prague","London","Europe"], keywords: ["European tech","startups","AI policy","tech ecosystem"], limit: 15)

clay - searchContacts (MCP)(query: "people in education, teaching, or academic integrity", job_title: ["Teacher","Lecturer","Professor","Educator","Academic","Curriculum Designer"], company_name: ["Czechitas"], location: [], keywords: ["education","teaching","AI detection","academic integrity","plagiarism"], limit: 10)

✻ Orbiting… (1m 38s · ↓ 557 tokens)

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24. 3. 9:08
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
claude
Claude-Vault | Opus 4.6 (1M context) | 3% | (main | 1 files +1 -1 ) [$0.27]
⌘K to generate command

[6:19](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=379s) [Terminal shows a command approval prompt for extracting contact information from search results.](https://assets.ai.engineer/sites/aiecode2025/sha256/d6e0c05668b9e2e4eae578df362628d90c0b0a39badd6731bc13251bbe88ed92.webp)

Read 2 files (ctrl+o to expand)

Bash(cat /Users/simon/.claude/projects/-Users-simon-Documents-Claude-Vault/9b42893a-c9c2-4091-87e7-b4e69caed745/tool-results/toolu_01XTLtJbLoNs8rLGeWq5G2YX.json | py…)
  Running…
Bash(cat /Users/simon/.claude/projects/-Users-simon-Documents-Claude-Vault/9b42893a-c9c2-4091-87e7-b4e69caed745/tool-results/toolu_01EhMKPNAUcqBUgS2AajXfGb.json | py…)
  Waiting…

Bash command
cat /Users/simon/.claude/projects/-Users-simon-Documents-Claude-Vault/9b42893a-c9c2-4091-87e7-b4e69caed745/tool-results/toolu_01XTLtJbLoNs8rLGeWq5G2YX.json | python3 -c "
import json, sys
data = json.load(sys.stdin)
contacts = json.loads(data[0]['text'])
for c in contacts:
    notes = c.get('notes', [])
    notes_str = '; '.join(notes[:2]) if notes else ''
    work = c.get('work_history', [])
    current = [w for w in work if w.get('is_active')]
    current_str = ', '.join([f\"{w['title']} @ {w['company']}\" for w in current[:2]]) if current else (f\"{work[0]['title']} @ {work[0]['company']}\" if work else '')
    print(f\"{c['name']} | {c.get('location','')} | {current_str} | {c.get('headline','')} | Notes: {notes_str}\")
"

Extract key contact info from AI/tech search results

Do you want to proceed?
❯ 1. Yes
  2. No

Esc to cancel · Tab to amend · ctrl+e to explain

Interface text:
Cursor  File  Edit  Selection  View  Go  Run  Terminal  Window  Help
Tue 24. 3. 9:09
Claude-Vault
Terminal  Problems  Output  Debug Console  Ports
claude
claude
⌘K to generate command

### In Practice: Cross-Source Query

An example combines browser sessions, notes, CRM and email to connect a week of reading about eval frameworks with three contacts, two of whom are going cold.

[7:14](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=434s) [A slide combines browser sessions, notes, CRM and email to identify relevant contacts.](https://assets.ai.engineer/sites/aiecode2025/sha256/5c802f807b740c9825dd2b78409cad616a225fa8fd4b0d68886cb06ef94f6025.webp)

In Practice: Cross-Source Query

Sources crossed:
• Browser sessions — what you’ve been reading
• Notes — what you’ve been thinking about
• CRM — who you know, when you last spoke
• Email — relationship recency and additional context

You've been reading about eval frameworks all week. Three people in your network work on this: A (last email: 2 months ago), B (met at a conference, never followed up), C (active collaborator). A and B are going cold.

No source knows all of this. Your browser doesn't know your contacts. Your CRM doesn't know what you're reading. The exhaust does.

### Why Read-Only?

Podhajský argues that read-only errors can be ignored, while write errors have potentially unbounded consequences for relationships, careers and reputations.

[7:26](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=446s) [A peach section slide asks why the system is read-only, beside a shield illustration.](https://assets.ai.engineer/sites/aiecode2025/sha256/82abfb0988ac8d9c5d9f3c034d1477ec1cdc46faa0cd223a53faba808df40b79.webp)

Why Read-Only?

### The Risk Table

Podhajský contrasts an irrelevant read-only suggestion he can ignore with a read-write error that can send a wrong email or delete a file, potentially requiring damage control.

[7:50](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=470s) [A table compares the consequences of read-only and read-write AI errors, followed by a preference for avoiding automated email misfires.](https://assets.ai.engineer/sites/aiecode2025/sha256/7a71130e7366e405ce0af6e85f69e7a138244b8de5d1721eee3fd5827ce7b611.webp)

The Risk Table (for AI mistakes — data breaches are a separate problem)

Observer (read-only) | Agent (read-write)

Best case
Observer: "I see a pattern I missed"
Agent: "It handled that task for me"

Worst case
Observer: Shows me something irrelevant
Agent: Sends wrong email, creates false commitment, deletes a file

Cost of error
Observer: I ignore it
Agent: Damage control, apology, lost trust

Recovery
Observer: Zero-cost
Agent: Potentially irreversible

I'd rather miss out on automated emails than have a misfire nuke my life.

### The Signal Purity Argument

Podhajský argues that AI edits contaminate the personal traces an observer analyzes. Reading a reflection and deciding what to do keeps the feedback loop mediated by the person.

[8:32](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=512s) [The slide argues that AI-generated changes can contaminate personal data and that unmediated LLM output can erode agency.](https://assets.ai.engineer/sites/aiecode2025/sha256/00f099fd43d06e102187c693f315d0bbf729a7203c0186f528f121baa6a288dd.webp)

The Signal Purity Argument

If the AI modifies your data, it's analyzing its own output.
Your email drafts, browser tabs, journal entries — these are authentic traces of your cognition. The moment an AI starts writing emails or organizing files, it pollutes the signal.

Handing your life over to an LLM reverts you to the mean.
Unmediated LLM output slowly erases your agency. Reclaim it.

### What This Means for Personal AI

Podhajský argues that a read-only observer serves a different need from task tools: it can reveal gaps between what you say matters and how you behave.

[9:24](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=564s) [The slide contrasts task tools with an observer that reveals personal patterns and describes read-only AI as a distinct product category.](https://assets.ai.engineer/sites/aiecode2025/sha256/b3a704f80eb1e81d264dbe0cad2a6d9c839ac659289690621a266db542e40de6.webp)

What This Means for Personal AI

You already have good tools for doing things.
Calendar, email client, task manager — they work fine. You have zero tools for understanding your own patterns.

The observer fills a gap agents don't try to address.
No amount of task automation gives you the insight: "you keep saying X matters but your behavior says otherwise."

Read-only isn't a stepping stone to "real" agents.
It's a different product category. The mirror and the butler serve different needs.

## Full transcript

[0:01](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=1s) Hi, my name is Šimon. Today I'll talk about a personal AI system that knows you but won't do anything instead of you or on your behalf, and won't blow up your life.

[0:13](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=13s) So that's good. In the process, I'll talk about the risks of personal AI and how read-only AI systems like this one mitigate them.

[0:21](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=21s) Let's get started. The whole personal AI space is obsessed with agents that act on your behalf. I built something different.

[0:30](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=30s) The starting point, six sources, read access only, no write permissions. The limitation is fully intentional.

[0:39](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=39s) But first of all, what are cognitive exhaust fumes? What are they? It's my term for the digital activity that is a byproduct of your cognition, like exhaust fumes for a car engine.

[0:50](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=50s) Individually, it's just waste, but if you analyze the exhaust, you can diagnose the engine.

[0:56](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=56s) So let's see some examples of what the exhaust reveals.

[1:00](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=60s) What does this enable you to do? Three top uses I found: intention-action gaps, attention drift,

[1:08](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=68s) and relationship decay. No single source tells you any of this, and the cross-source ability is what these have in common. Your email client doesn't know what you journaled. Your task manager doesn't know what you're browsing.

[1:23](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=83s) The cross-source signal is the product. Let's take a closer look at the system.

[1:29](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=89s) Here it is, full in. Three zones. The sources are read-only. The AI never writes back to them. The workspace is where the analysis happens. The outputs land in a separate Obsidian vault for me to review, but it doesn't have to be a separate Obsidian vault.

[1:45](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=105s) It could be a separate Notion, separate text file, separate anything. Could be any other system. And that's the whole thing.

[1:54](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=114s) So what about applications? Let's start with a David Allen style Getting Things Done like spin on the weekly reflection. Based on the six sources, the AI synthesizes an occasionally brutal reflection on how you spent your week.

[2:07](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=127s) Let's look at a real example. Everything runs in Claude. I've stored this logic in the weekly reflection slash command slash skill. And what it does is it launches a Python script that gets all the data that come from the read-only sources and looks through them and creates structured outputs with

[2:32](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=152s) specified prompts that I've prepared. This takes a little bit of a while. It pings the Anthropic API to get those structured outputs back. And once it does, it will create a Markdown document that I will be able to review.

[2:51](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=171s) This is now all finished running, so it gives me an overview, and I can open it back up in Cursor and I'll convert it to a preview that's more readable and see that in fact, it does hit the themes of the week.

[3:09](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=189s) It does hit some of the tensions and conflicts that I need to think about.

[3:14](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=194s) Talks about my commitments and relationships, which is mostly notable by its omissions, and highlights the notable moments as well as reflection questions that I like to think about.

[3:31](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=211s) In short, this is not a productivity report. It's a reflection on how you're thinking, assembled entirely from exhaust.

[3:38](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=218s) Let's take another example. I like to discuss what I'm reading with others, but sometimes I think I shouldn't keep messaging the same three people about it. So I ask the AI, given my recent reading, who in my network should I be discussing this with?

[3:53](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=233s) This is the cross-source magic. Four data sources, none of which were designed to talk to each other, combined into an insight you'd never get from any single tool, and all read-only.

[4:03](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=243s) Nothing was sent, nothing was scheduled. Just a suggestion for me to act on if I choose.

[4:10](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=250s) Let me once again show you the demo.

[4:14](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=254s) Once again, this is a Claude skill, but in this case, I've hidden most of the pa- guts of the Claude skill into the cross-source query and ask for the specific question, uh, in plain language.

[4:30](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=270s) The plain language knows that it will auto, uh, that it will activate the specific skill for the cross-source queries, and it now goes through the databases that I've curated, that I have regular ingest for, looks through the Vivaldi SQLite database for the articles that I've been reading the most.

[4:52](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=292s) And after a while, it will figure out which of these articles, um, are most read, still open on tabs, and which people might be curious about it based on the profile.

[5:08](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=308s) Now, this is probably the weakest part. The Clay MCP takes forever to run, but it searches my CRM or my friend relationship system, I suppose so, FRM, for people who might be interested in articles on, on these topics.

[5:27](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=327s) Um, in this case, it's people interested in AI or people in European tech or people in education, which coincidentally are three things that I also am.

[5:42](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=342s) Now, as you might notice, this takes up a lot of tokens in the context window, so you probably don't want to do this in a session that is not clean, but it's not a problem if it, uh, messes up a little bit of the one million context window for 4.6, and then you clear it again.

[6:02](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=362s) So at this point, it's getting the responses from all of the Clay searches. It synthesizes the people that I should talk to, and it maps them to one article each.

[6:16](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=376s) That's what the... Or it's about to map them to one article each. Uh, this requires a little bit of, uh,

[6:25](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=385s) bash sorcery on behalf of Claude Code, but if you run it with auto to auto mode or --dangerously-skip-permissions, you can get rid of that as well.

[6:34](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=394s) And indeed, when I take a look at the first res- results, those look like the people that I might want to talk to that I haven't talked to yet about the kinds of articles that I've been reading.

[6:46](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=406s) So thank you, Claude. Right. In this case, it even found the author of the article that I was reading that's in my network, so I should give them a whirl.

[7:13](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=433s) In short, no source knows all of this. Your browser doesn't know your contacts. Your CRM doesn't know what you're reading. The exhaust does.

[7:28](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=448s) So why keep it read-only if it's so useful after all? Here's the thing about the risk involved. It's asymmetric. The downside of a read-only error is zero. I just ignore it.

[7:39](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=459s) The downside of a write error is unbounded, and personal AI operates in the highest stakes environment: your relationships, your career, your reputation.

[7:51](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=471s) I'd rather miss out on automated emails than have a misfire nuke my life.

[7:57](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=477s) But there's also a subtler philosophical argument, almost a matter of taste. Read-only isn't just safer, it produces better analysis. The moment your AI writes to your data sources, the exhaust fumes are contaminated.

[8:10](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=490s) You're no longer observing your cognition. You're observing a human AI hybrid, and you can't tell which patterns are yours. Sure, the observer changes your behavior too, but the feedback loop is mediated by you, not automated.

[8:25](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=505s) You read the reflection. You decide what to do. That's a different thing from the AI rewriting your draft. And there's an argument to be made that you don't want the AI to write your draft in the first place, that you should reclaim your agency.

[8:39](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=519s) That might be a hard sell for this crowd, I think, but worth considering.

[8:44](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=524s) At this point, you might be asking, why not throw all this into OpenClaw on a read-only mount? Which I have.

[8:51](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=531s) Here's the thing. The produce-- the observer produces more value per interaction by a wide margin. The agent saves me thirty seconds on a weather check. The observer shows me I've been avoiding my most important project for two weeks.

[9:06](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=546s) Not to mention that there's less risk of exfiltration and cognitive pollution.

[9:13](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=553s) The argument I'm making here is that read-only isn't a stepping stone to, quote unquote, real agents. It helps you do things well, yes, but it fits a different gap, serves a different need.

[9:24](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=564s) It's a different product category. The industry frames read-only as a limitation you graduate from. I think that's wrong. Observers and agents are different tools. A mirror isn't a broken butler.

[9:40](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=580s) So that's the value proposition. But I'd be doing you a disservice if I stopped here. Let's put on the paranoid hat.

[9:50](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=590s) What keeps me up at night? Let's start with the mosaic effect.

[9:55](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=595s) There's something called the mosaic effect, where you put together a lot of small pieces of information and you get a picture.

[10:03](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=603s) My own slide copy describes the security risk perfectly. The same cross-referencing that makes the system useful makes it a devastating target.

[10:11](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=611s) So careful there. The other side of the coin, Simon Willison's lethal trifecta. In case you don't know the lethal trifecta, it's a security risk model that combines three factors: private data, untrusted content, and external communications.

[10:29](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=629s) I initially thought read-only broke the lethal trifecta, and it doesn't, not fully. It removes the natural exfiltration cir-- uh, channels, but the third leg is any ability to communicate externally, and shell access still has that.

[10:45](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=645s) In short, the system isn't fireproof, and I'm not claiming that. Even in the best case scenario, I'm still sending data to Anthropic on a network that's mostly open with a lot more information lying around than is strictly speaking required.

[11:00](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=660s) I'm not claiming the system is secure. I'm claiming that I've thought about where it isn't, and I've decided which risks I'm willing to carry. It's different from not knowing.

[11:09](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=669s) The worst security posture is the one you haven't examined.

[11:14](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=674s) With that said, I still think there's something worthwhile to be learned from all this. Your digital exhaust is the most underused data set you own. Reflect on it and use it to make yourself better.

[11:26](https://www.youtube.com/watch?v=u0TOSBbAw7c&t=686s) Thanks for listening.
