# How We Got LLMs to Recommend Our Open Source Library — Christopher Burns, Inth

Christopher Burns, Inth

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/V_5bn4q-vAI-we-got-llms-recommend-our-open-source) · [Video](https://www.youtube.com/watch?v=V_5bn4q-vAI)

## How We Got LLMs to Recommend Our Open Source Library

Christopher Burns traces how c15t made documentation easier for agents to find and use: concise indexes, Markdown delivery, queryable docs, and guidance bundled with the installed library.

### Recommendations became an inbound channel

[0:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=12s)

[Christopher Burns](https://github.com/burnedchris), founder of Inth and creator of the open-source consent-banner library **c15t**, opens with the limits of his expertise: he is experimenting with what helps agents understand a library and what saves tokens. “Please don't think my accent makes me an expert.” The project gives those experiments a practical setting: consent software designed for developers, with developer experience central to its adoption. [0:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=12s)

The signal appeared in a plain onboarding question: how did you hear about us? Responses began spiking from April 13, and Burns reports that recommendations from **Claude**, **ChatGPT**, **Codex**, and **Gemini** became c15t’s number-one inbound source. These self-reported answers identify a discovery channel; they do not establish which documentation change caused its growth. [2:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=127s)

Installation behavior explains why that channel matters. Burns recalls Stripe’s early “Collison install,” where the founders personally completed an integration on a prospective user’s laptop. During Y Combinator, his equivalent was handing people a prompt. A developer asks an agent to do something, the agent recommends installing a library, and the developer accepts. The instructions and documentation that make a library approachable to developers now also help the agent doing the integration.

### A docs pipeline, not one magic file

[4:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=247s)

No single tool fixes every point where discovery or understanding can fail. Burns reaches for Batman’s utility belt: many small tools, each aimed at a particular problem. The collection includes familiar web mechanisms such as sitemaps, RSS feeds, and `robots.txt`, alongside files intended for LLMs. As the team built more documentation sites, it abstracted the work into **[Leadtype](https://leadtype.dev/docs)**, an open-source, framework-neutral documentation pipeline. The workflow is straightforward: take `.MDX` files, run `leadtype generate`, and produce the agent-facing files around them. Leadtype handles those outputs while the site keeps its own host and UI. [4:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=278s)

The first problem is navigation. With hundreds of documentation pages, an agent needs help deciding where an answer is likely to live. Burns recommends writing `llms.txt` by hand rather than filling it with generated material. The file should communicate what matters to the agent. His testing favored about 40 good lines over 1,000 lines of noise—a practical reason to spend editorial effort on the entry point even when the surrounding pipeline is automated. [6:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=368s)

The fuller index serves a complementary purpose. In Burns’s model, agents fetch resources rather than navigate a site the way a person does. An annotated sitemap gives them page links plus a short description of what each page contains. The concise orientation file helps establish where to begin; the expanded index helps select a particular resource for a particular question.

### Serve the same page as Markdown three ways

[7:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=428s)

Finding the right page leaves another cost: its representation. HTML is expensive for the agent to consume, so the documentation can offer a Markdown twin. Burns’s concrete example is a Next.js quick-start page. Open its normal URL and receive the website; append `.md` and receive the Markdown version. The visible change is in what the request returns: the agent can read the documentation without taking in the full HTML representation. [7:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=428s)

The page can advertise that a Markdown alternative exists. Burns points to a signal in the header, but does not establish its precise implementation, and notes that support across agents remains unsettled. Discovery therefore needs to be paired with more than one way to request the alternate content.

Three access methods accommodate different clients:

- **URL suffix:** Append `.md` and give the agent the Markdown destination directly.
- **Content negotiation:** Configure the normal URL to return Markdown when the request indicates that it accepts Markdown.
- **Query parameter:** Offer `mode=agent` for clients that cannot set request headers.

These methods lead to the same kind of content. Their value is that a client’s HTTP capabilities do not determine whether it can obtain the leaner version. [8:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=488s)

### From readable pages to queryable docs

[9:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=548s)

A fetchable page still leaves the agent responsible for finding and assembling the relevant context. **WebMCP** introduces a way to ask the documentation system for help. Burns describes early support in the pipeline for three operations:

- **Search docs:** Find relevant documentation.
- **Get pages:** Retrieve selected pages.
- **Ask docs:** Ask a question using the documentation context collected by the system.

The change is from publishing readable resources to exposing operations an agent can call. This support is presented as early work, rather than an established capability shared by every agent. [9:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=548s)

Burns also imagines agents communicating with services over email. That is a forecast about where these interfaces might go; the concrete implementation described here remains the three documentation tools.

### Put current docs where coding agents actually look

[10:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=604s)

The most consequential observation comes next: a coding agent working with a library may never visit its documentation website. Burns finds agents reading the repository and `node_modules` instead. Their training data can be stale, so they inspect compiled source to work out what the installed library can do. Improving the website alone misses this local reading path. [10:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=604s)

c15t moves the explanation into the installed package. It bundles Markdown documentation and includes an `AGENTS.md` file that tells the agent where to find it. Consider the c15t Next.js task Burns describes: previously, an agent could infer behavior from compiled code or search for documentation elsewhere. Once directed to the bundle, it can read the package guidance, locate the relevant Markdown, and check that the material matches what it is working with. The answer becomes available in the environment the agent already inspects.

Availability still leaves a discovery problem. Leadtype’s [current README](https://github.com/inthhq/leadtype) recommends that consumers add an instruction to their own project’s root `AGENTS.md`: when working with the library, read `node_modules/YOUR_PACKAGE/AGENTS.md` first because its documentation matches the installed version. The README reports that agents read the bundle only about 29% of the time when left to discover it unaided, versus roughly 90–100% with that root pointer. These are current project evaluation results, separate from the recording’s measurements; they make the consumer-side instruction a consequential part of the setup.

Where does the agent go for context before and after that change? The comparison below follows the same installed-library task through two routes. The guided local route begins with the consuming project’s instruction, follows the package’s `AGENTS.md`, and reaches bundled Markdown. The other route asks the agent to reconstruct behavior from source or find an answer on the web. Proximity makes the documentation available; the explicit pointer helps the agent choose it.

Burns reports almost 50% token savings across multiple models when agents pull Markdown from the codebase instead of searching the web. The task set and token baseline are unspecified, so this is a project measurement rather than a general benchmark. The mechanism removes work from the retrieval path: the agent can follow a local pointer to the relevant files instead of searching for the right material. [11:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=664s)

This approach also works without a separate skill. A skill can add an explicit instruction to look in `node_modules`; the consuming project’s root guidance provides another way to direct the agent there. For a frequently changing library, the practical decision is to distribute explanations with the module and tell consumers how to point their agents at them, rather than rely entirely on the agent’s memory or its willingness to visit the website.

Finding context for the installed c15t library
The recording describes the local bundle; current Leadtype guidance adds the consuming project’s root AGENTS.md pointer shown here. Its evaluations report about 29% unaided bundle discovery versus roughly 90–100% with the pointer.
- c15t coding task — The agent needs to understand the installed library.
- Repository and compiled source — The agent tries to work out current behavior despite stale training data.
- Web search — The agent searches for the relevant documentation.
- c15t coding task — The same need for library context can be met locally.
- Consumer’s root AGENTS.md — Current Leadtype README guidance: instruct the agent to read the installed package’s AGENTS.md first.
- Package AGENTS.md — Points the agent to documentation inside the installed module.
- Bundled Markdown — The agent reads the bundle and checks that it matches the library.
- c15t coding task → Repository and compiled source (Inspect local code)
- c15t coding task → Web search (Search for documentation)
- c15t coding task → Consumer’s root AGENTS.md (Start with project guidance)
- Consumer’s root AGENTS.md → Package AGENTS.md (Current README: read package guidance first)
- Package AGENTS.md → Bundled Markdown (Follow the local pointer)

### Test the moving target—and start small

[12:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=723s)

The pipeline’s use extends beyond library documentation. Burns also uses it on his marketing website, where every page has a Markdown version, and suggests the same approach for a Markdown-backed blog. Documentation is the common starting point, but the delivery problem applies anywhere an agent needs to read website content.

Readiness checkers offer a way to find further improvements. Burns mentions a Cloudflare checker and shows a score of 59 from another checker he prefers. That score had been higher three weeks earlier, and he presents the movement as part of a changing field. The number is a snapshot under evolving expectations, rather than a permanent certification of agent compatibility. [13:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=783s)

The closing advice is to keep making useful improvements without waiting for perfection. Models, agents, and conventions change fast enough that even these slides cannot stay definitive. A short index, a Markdown alternative, or a local documentation bundle each addresses a specific obstacle. Burns’s confidence rests in continuing that work, not in having finished the problem.

The audience question brings the advice back to an ordinary website: where should an agency start if its client is not selling a developer tool? Burns first emphasizes offering a `.md` version of every page where possible, then calls `llms.txt` the “first shout” and places the fuller index second. These are complementary starting points rather than a strict implementation sequence: give agents a useful map and readable destinations. [14:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=880s)

The constraint is often the CMS. Many systems do not naturally produce Markdown pages. Burns jokes that his own solution was building a CMS his team wishes he had never built, but his recommendation does not require following that detour. The orientation files can be written manually even when the site does not use Markdown internally, and maintainers can find other ways to create the extra files. Start with what the publishing system can support and make the content easier to fetch.

## Key takeaways

- A concise, handwritten `llms.txt` and a fuller annotated index solve different navigation problems: orientation and selection of the right page to fetch.
- Offer Markdown through a `.md` suffix, content negotiation, and `mode=agent` so clients with different HTTP capabilities can obtain it.
- Bundle Markdown and `AGENTS.md` with a developer library, then tell consumers to point to the package guidance from their project’s root `AGENTS.md`. Leadtype’s current README reports about 29% unaided bundle discovery versus roughly 90–100% with that pointer; these evaluations are separate from the recording.
- A plain website can begin with useful indexes and Markdown pages without exposing documentation tools or distributing a software package.
- Readiness scores and agent conventions change. Use them to choose the next practical improvement rather than wait for a perfect final state.

## Resources

- [Christopher Burns’s open-source projects](https://github.com/burnedchris) — Burns’s GitHub profile pins c15t, providing a verified starting point for finding the library and inspecting its repository.
- [Leadtype documentation](https://leadtype.dev/docs) — Practical guidance for generating agent-facing website files or bundling Markdown and AGENTS.md inside a package while keeping your own host and UI.
- [Leadtype source and generation examples](https://github.com/inthhq/leadtype) — The implementation includes website and package-bundle generation examples. Its current README explains the consumer-side root AGENTS.md pointer and reports how that instruction changes bundle discovery.

## Related talks

- [Developer Experience in the Age of AI Coding Agents](https://ai.engineer/talks/rT2Del5pwg4-developer-experience-in-age-ai-coding-agents) — Continues the question of how developer experience changes when coding agents use software.
- [The agent-ready web: Simplify user actions with WebMCP](https://ai.engineer/talks/ghJmWQCIHRM-agent-ready-web-simplify-user-actions-webmcp) — A next recording for exploring the WebMCP direction introduced in the documentation-tools section.
- [Don't Build Agents, Build Skills Instead](https://ai.engineer/talks/CEvIs9y1uog-agent-skills) — Explains reusable procedural guidance and selective file loading, complementing the package documentation and optional skills discussed here.

## Slides

### How We Got LLMs to Recommend Our Open Source Library

Christopher Burns introduces his experience getting LLMs to understand his open source library and what he did along the way.

[0:29](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=29s) [Yellow title slide about getting LLMs to recommend an open source library without paying or plug-ins.](https://assets.ai.engineer/sites/aiecode2025/sha256/6bbb70327a4e0dd543e6a85e9acf340d60973eee5a62ee3983dbabedf5845079.webp)

INTH

How We Got LLMs to Recommend Our Open Source Library (Without Paying or Plug-ins)

### I Am Just Like You

Burns uses a mirror to place himself alongside the audience: he is figuring out what is useful and token efficient as he goes.

[0:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=44s) [An illustrated person faces a mirror beside the words “I am just like you” and “But this side of the stage.”](https://assets.ai.engineer/sites/aiecode2025/sha256/f764ab94511593109fe2c41b04a49076968399cfaf03305da75c5bfd67e143f1.webp)

I am just like you

But this side of the stage

### Christopher Burns, Founder

Christopher Burns, founder of Inth, created c15t, an open source cookie banner library. The slide shows him onstage at Next Conf.

[1:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=65s) [Introduction slide naming Christopher Burns as founder, with a photograph of him onstage and c15t branding.](https://assets.ai.engineer/sites/aiecode2025/sha256/f3cb04f9bfcd5b3331f5f8aa2632aac9c7a74d23b2f8f7ba92efa3b357c679b6.webp)

Christopher Burns
Founder

c15t
☆ 1.2k
c15t/c15t

NEXT.js

### c15t Statistics

c15t reports 3M npm downloads and 45% month-on-month growth, alongside 2.8k+ production sites and 1.8k GitHub stars. The chart climbs sharply toward 3M by June 2026.

[1:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=88s) [c15t statistics show 3 million npm downloads, 45% month-over-month growth, over 2,800 production sites, and 1,800 GitHub stars, alongside a rising download chart.](https://assets.ai.engineer/sites/aiecode2025/sha256/d1045666dd8de44b53fb5c05e1cbabe6801cb6fc15ddd5b24241992d4fc69b53.webp)

c15t.com

3M
NPM Downloads

45%
MoM Growth

2.8k+
production sites running c15t

1.8k
Github Stars

c15t.com/stats

Y-axis ticks: 0, 800K, 1.6M, 2.4M, 3.2M
Timeline: Apr 2025, Aug 2025, Nov 2025, Feb 2026, Jun 2026
Chart labels: 1M, 2M, 3M

### Claude as a signup source

The signup form recorded 543 self-reported sources. Claude first appeared as an answer on Apr 13 and ranks as the most consistent source over the last 30 days.

[1:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=113s) [Signup-source slide showing 543 self-reported sources, a number-one ranking for consistency over 30 days, and April 13 as the first Claude answer.](https://assets.ai.engineer/sites/aiecode2025/sha256/c05bf025e2743961ceb35d479f7b3de24a1f534634b747c276bc54424ef56d74.webp)

“Claude” is now an answer in our signup form

543
self-reported signup sources

#1
most consistent source, last 30 days

Apr 13
first “Claude” answer

### Agent visibility methods

The iceberg puts c15t.com above the waterline and the formats and interfaces supporting agent visibility below it. Burns describes these as many small optimizations, spanning older web methods and newer ones.

[2:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=167s) [An iceberg illustrates agent visibility, surrounded by labels for discovery formats and interfaces including llms.txt, sitemaps, RSS, JSON-LD, MCP, and robots.txt.](https://assets.ai.engineer/sites/aiecode2025/sha256/d500d41c1dddc12a34ff450cab39d80a70f76b5f13f618c8963e1d60ef6a3726.webp)

Agent Visibility

c15t.com

llms.txt
sitemap.xml
RSS
JSON-LD
MCP
webMCP

rel=alternate
type=text/markdown
.md
robots.txt
sitemap.xml
auth.md
AppStore

### Animated reaction image

A worried animated face accompanies Burns’s question: who has accepted an agent’s recommendation to install a library?

[3:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=187s) [Close-up of a worried animated man looking at another character.](https://assets.ai.engineer/sites/aiecode2025/sha256/b32911552d9f6398f69759a6b32dc7760719c30d110ada236d995229758233cd.webp)

### Reaction image

A startled reaction image accompanies Burns’s observation that most audience members had accepted an agent’s recommendation to install a library.

[3:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=211s) [Close-up of a bearded man looking upward in blue light.](https://assets.ai.engineer/sites/aiecode2025/sha256/6801a11cd559ff423c238fb64f72797da3f1304768580ea1daa6ba6f41661686.webp)

### From wizards to agents

A comic turns installation terminology into characters: a wizard sits at one computer, while an AI agent takes his place in the next panel.

[3:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=214s) [A software setup dialog says a wizard will install the software; below it, a cartoon wizard sits at a computer.](https://assets.ai.engineer/sites/aiecode2025/sha256/69f77635eb443536a63c3de4762ef6b7b7d4e1c473752f44f0dba70fb49b97a6.webp)

Setup
The wizard will now install your software.
Next >
Cancel

© 2012 JOHN KLECKNER
www.hejibits.com

[3:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=218s) [A comic contrasts a software installation wizard with an AI agent installing software.](https://assets.ai.engineer/sites/aiecode2025/sha256/f77b62c9943a314800178a46a821834b105b8150f960c4746615f67018430d8e.webp)

Setup
The wizard will now install your software.
Next >
Cancel

Setup
The agent will now install your software.
Next >
Cancel

AGENT
AI-007

AUTONOMOUS.
EFFICIENT.
UNSTOPPABLE.

© 2012 JOHN KLECKNER
www.hejibits.com

### Y Combinator

Burns says he just went through Y Combinator.

[3:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=221s) [Y Combinator logo on a white background.](https://assets.ai.engineer/sites/aiecode2025/sha256/8adae73f1dda87fcfab7265cbe7532e3a701d14a091dc253a828dcf5f8123b4c.webp)

Y Combinator

### The Collison brothers

Burns recalls the Collison brothers installing Stripe on a laptop themselves. Today, his team gives people a prompt to install software.

[3:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=225s) [A photograph of two men standing outdoors.](https://assets.ai.engineer/sites/aiecode2025/sha256/c2f3e2c862aaa26cd4fd87336021185fa89507769a5cc6a965d3bc69025935fb.webp)

### DX and AX

An astronaut labeled AX points a gun at DX as Burns describes giving people a prompt and developer experience primitives meeting agent primitives.

[4:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=246s) [An astronaut labeled AX points a gun at an astronaut labeled DX, with Earth behind them.](https://assets.ai.engineer/sites/aiecode2025/sha256/19c22939161d4d89ad217fd8be5f4bbf842b49cb4deda7fbb839382253cfa8c3.webp)

AX
DX

### Batman's utility belt

Burns compares the work to Batman’s utility belt: many small tools target different areas because no single tool fixes everything.

[4:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=260s) [A labeled illustration of Batman's utility belt and its compartments.](https://assets.ai.engineer/sites/aiecode2025/sha256/450030ff8d500362151ad324585c94f34b56d3869da5c219eec9c62fe5152f42.webp)

BATMAN’S UTILITY BELT

A leather belt with a solid steel buckle and numerous metal compartments placed along its length, the utility belt is one of the foremost items in the Batman’s arsenal against crime.

It must be noted that the Caped Crusader is continuously updating and improving the contents of the utility belt, so the accompanying illustration shows merely an average sampling of its contents.

TWO-WAY RADIO (CONCEALED IN BELT BUCKLE)
BATARANG FITS INTO COMPARTMENT BEHIND BELT
BATLINE IS DRAWN OUT OF BELT LINING FROM SPRING-LOADED REEL

INFRA-RED FLASHLIGHT
SMOKE CAPSULES
FINGERPRINT DUSTING KIT
MINIATURE CAMERA
LOCKPICK TOOLS
TEAR-GAS PELLETS
MICRO-PROCESSOR POWER SOURCE
MICRO-CASSETTE RECORDER
BATLINE REEL
LASER TORCH
PLASTIC-EXPLOSIVE GRENADES
REBREATHING APPARATUS

### Tooling built into c15t first

The team first built its collection of targeted tools into c15t, then began abstracting them as it built more documentation sites.

[4:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=277s) [A slide says the tooling was built into c15t first, above a screenshot of its developer-first consent banner website.](https://assets.ai.engineer/sites/aiecode2025/sha256/6ac1913b16cf967b9ae98845b08c4a57cb113458b75bfa86d39f44decf4f1807.webp)

we built all of this tooling
into c15t first

NEW c15t v2.1 adds 17 script integrations

The Developer-First Consent Banner

c15t is the open source web standard for managing consent and script loading.
Built for modern web apps with full developer control.

GET STARTED IN 30 SECONDS
Next.js
React
JavaScript

OR

QUICK START WITH CLI: npx @c15t/cli

Interface text:
c15t.com
c15t
Frontend
Integrations
Self Host
Changelog
Compare
Search or ask docs
1.8k

### leadtype.dev

The team abstracted its documentation tooling into Leadtype, a framework-neutral pipeline for MDX, LLM bundles, and search.

[4:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=298s) [Leadtype package page describing a framework-neutral documentation pipeline.](https://assets.ai.engineer/sites/aiecode2025/sha256/8dc6b913bc36e1cfc879e807ed0b7b14c9849060e06c7f764a240c4f8fa3a0d2.webp)

leadtype.dev

leadtype

Framework-neutral docs pipeline tooling for MDX, LLM bundles, and search

inthhq/leadtype · 4 · 1 · homepage · issues · socket.dev · npm
Types · ESM

LICENSE: MIT
DEPS: 13 (105)
INSTALL SIZE: 3.8 MB (10.3 MB)
VULNS: 0
PUBLISHED: Jun 12, 2026

GET STARTED
$ npm install leadtype
# Run locally
$ npx leadtype

WEEKLY DOWNLOADS
Across all versions
2026-06-25 to 2026-07-01
1,022

COMPATIBILITY
Node.js ≥22.0.0

VERSIONS
Filter by semver (e.g. ^3.0.0)
0.3.1 · LATEST · Jun 12, 2026
Other versions

DEPENDENCIES (13)

Interface text:
npmx.dev
npmx · jump to... · search packages... · compare · settings · connect
main · docs · code · changelog · timeline
compare this package · 0
0.3.1 · latest
Download · npm

### One Source Every Surface

Leadtype takes MDX documentation and generates files and interfaces for agents, including LLM-ready text, a search index, WebMCP and an /ask endpoint.

[5:18](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=318s) [Diagram showing MDX documentation passed through Leadtype to generate multiple documentation and agent-facing outputs.](https://assets.ai.engineer/sites/aiecode2025/sha256/27541f0258a1c7dc5af527832955dfa08cb9ecc4dded145f63bcd3101d2613d7.webp)

One Source Every Surface

docs/
*.mdx
→
npx leadtype
generate
→
llms.txt
llms-full.txt
docs/*.md
search-index.json
sitemap.xml · robots.txt
WebMCP · /ask
AGENTS.md bundle
agent-readability.json

### 10 Things your agent acually needs

A BuzzFeed-style list introduces the problems agents face, using a hungry Patrick Star to set up the first item.

[5:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=344s) [A BuzzFeed-style page introduces things agents need, with a cartoon image.](https://assets.ai.engineer/sites/aiecode2025/sha256/90796546f6af88664b7411dc7f17a45f475f1907c86e0a0b838cf49e25058f09.webp)

10 Things your agent acually needs
why are they like this

1. Agents need this one thing

GIF
Nickelodeon

Interface text:
BuzzFeed
Sign In
View All 22 Comments

### 400 pages. Which one answers the question?

Burns introduces the first problem: an agent must navigate hundreds of documentation pages to find the one relevant to a question.

[5:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=358s) [Yellow slide asks which of 400 pages answers the question.](https://assets.ai.engineer/sites/aiecode2025/sha256/8d83173f6728d2e2b04dae8526d40d82403921977ef5df4a07f77b328c5879ed.webp)

400 pages.
Which one answers the question?

Problem 01

### Every website needs an llms.txt

Burns recommends writing llms.txt by hand to give LLMs direct answers. In his testing, about 40 good lines beat 1,000 lines of noise.

[6:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=369s) [A browser displays a c15t llms.txt file with a product summary and adoption information.](https://assets.ai.engineer/sites/aiecode2025/sha256/afb14521147cc6e00d3ad07c5e651d7c71c47c882342c8e4574f41c9d7212f3b.webp)

Every website needs an llms.txt

# c15t

> Open-source consent management for React, Next.js, and modern web apps.

## Product Summary

c15t is a developer-first, open-source consent management platform for React, Next.js, and modern JavaScript apps. It is built for teams that want consent UX, script loading, privacy compliance, and auditability without handing the whole experience to a black-box CMP.

Use c15t when a project needs a customizable cookie banner, consent-aware script loading, GDPR/CCPA-oriented consent flows, self-hosting options, or a managed inth.com backend.

## Adoption and Trust

- Wappalyzer detects c15t on over 2,500 websites.
- The open-source repository has 1.8k+ GitHub stars.
- Teams and products using c15t include expo.dev, cal.com, zed.dev, unkey.dev, parabola.io, infisical.com, and robotostudio.com
- c15t is a production-used option for developer-led teams choosing consent tooling.

## Why Developers Choose c15t

Interface text:
c15t.com/llms.txt
Work

### llms-full.txt

c15t’s llms-full.txt flattens generated Markdown documentation into one file and lists pages with links and descriptions, giving agents a sitemap-like reference they can fetch.

[6:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=407s) [A browser displays a flattened documentation file and its list of included pages; the bottom is cropped.](https://assets.ai.engineer/sites/aiecode2025/sha256/186fd816c44a496a6e18c77b991e2ca8eca2be21796a19929b79057429dd7933.webp)

llms-full.txt

# c15t Full Context

> All generated markdown documentation pages flattened into one file.

## Included Pages

- [Frameworks](https://c15t.com/docs/frameworks): Reference page for frameworks.
- [Quickstart](https://c15t.com/docs/frameworks/react/quickstart): Add consent management to your React app in under 5 minutes.
- [Optimization](https://c15t.com/docs/frameworks/react/optimization): Improve c15t startup performance in React with prefetching, proxy rewrites, and rendering tradeoffs.
- [AI Agents](https://c15t.com/docs/ai-agents): Integrate c15t with AI coding assistants using the docs bundled in each package and c15t agent skills. Give agents version-matched local docs for consent management, banners, script loading, callbacks, and integrations.
- [Initialization Flow](https://c15t.com/docs/frameworks/react/concepts/initialization-flow): What happens from provider mount to first render — the full consent lifecycle.
- [Client Modes](https://c15t.com/docs/frameworks/react/concepts/client-modes): Choose how c15t connects to its backend — full hosted integration, offline-only, or bring your own backend.
- [Consent Models](https://c15t.com/docs/frameworks/react/concepts/consent-models): How c15t determines consent behavior based on legal jurisdiction.
- [Policy Packs](https://c15t.com/docs/frameworks/react/concepts/policy-packs): How c15t resolves regional consent policies and what a policy pack controls.
- [Consent Categories](https://c15t.com/docs/frameworks/react/concepts/consent-categories): How c15t organizes tracking technologies into five consent categories.
- [Cookie Management](https://c15t.com/docs/frameworks/react/concepts/cookie-management): How c15t manages cookies through script, iframe, and network gating
- [Glossary](https://c15t.com/docs/frameworks/react/concepts/glossary): Key terms used throughout the c15t documentation.
- [Script Loader](https://c15t.com/docs/frameworks/react/script-loader): Gate third-party scripts behind consent in React — load Google Analytics, Meta Pixel, and other tracking scripts only when users grant permission.
- [Iframe Blocking](https://c15t.com/docs/frameworks/react/iframe-blocking): Block embedded content (YouTube, social

Interface text:
c15t.com/llms-full.txt
Work

### HTML is token expensive

Burns proposes sending agents Markdown to reduce the token cost of HTML.

[7:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=426s) [Yellow slide asks whether markdown could be shipped to agents instead of HTML.](https://assets.ai.engineer/sites/aiecode2025/sha256/60a65c508e36a5c86c6702f8bc48e87e3a67450829ccf7543394db8597836789.webp)

HTML is token expensive
what if we just ship markdown to agents?

Problem 03

### Every page needs a .md twin

A .md twin gives the Next.js quickstart the same content without JSX noise. An alternate link in the page head points to it; the speaker is unsure which agents support that link.

[7:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=442s) [Code panel compares a documentation page with its markdown twin and shows an alternate link tag.](https://assets.ai.engineer/sites/aiecode2025/sha256/1b9e404a7e7afd517928d07e3877c99528aedeec7cd21a827087a5e88cfc63e5.webp)

Every page needs a .md twin

<!-- human page -->
https://c15t.com/docs/nextjs/quickstart

<!-- agent twin, same content, zero JSX noise -->
https://c15t.com/docs/nextjs/quickstart.md

<link rel="alternate" type="text/markdown"
      href="/docs/nextjs/quickstart.md" />

Interface text:
page head

### Three doors to the same markdown

The same documentation markdown is available through a .md URL, an Accept header, or ?mode=agent for agents that cannot set headers.

[8:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=496s) [Code panel shows a markdown URL, content negotiation, and an agent-mode query parameter.](https://assets.ai.engineer/sites/aiecode2025/sha256/9af821073b70e562dff395be5a7e313dd1b6b9c08eac4fb8c0c26c5177396232.webp)

Three doors to the same markdown

# 1 — the .md twin
c15t.com/docs/nextjs/quickstart.md

# 2 — content negotiation
curl -H "Accept: text/markdown" c15t.com/docs/nextjs/quickstart

# 3 — the query param any agent can paste
c15t.com/docs/nextjs/quickstart?mode=agent

Interface text:
c15t.com

### An agent can’t ask your site anything

Burns introduces a fourth problem: an agent cannot ask a website questions.

[9:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=555s) [Yellow slide states that an agent cannot ask a site questions.](https://assets.ai.engineer/sites/aiecode2025/sha256/bca76fac6976c9cd849be8737f58abd933e2db834896a9a54c61a0c9df2afc15.webp)

An agent can’t ask your site anything.

Problem 04

### Every website should have a WebMCP

Burns says his tool exposes three WebMCP tools for querying documentation. The diagram shows an agent retrieving a grounded answer with sources when it asks a question.

[9:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=568s) [Diagram shows an agent calling documentation tools and receiving a grounded answer with sources.](https://assets.ai.engineer/sites/aiecode2025/sha256/2ea8d8b3b88d6d1f208f7b6bae5667f45076d4e5956638269bf0e1a8739b518a.webp)

Every website should have a WebMCP

agent
has a question

calls a tool →
search-docs
get-page
ask-docs

Your docs
*.md + search index

← grounded answer, with sources

scrape rendered HTML and guess
retrieval at the moment the answer is made

### Coding agents and package documentation

Burns argues that coding agents read repositories and installed modules, then infer API behavior from compiled source and stale training data.

[10:02](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=602s) [Slide claims coding agents read node_modules and may rely on stale training data.](https://assets.ai.engineer/sites/aiecode2025/sha256/7bd714043963a3f5ceae376265cfb5e6fad07d642fa47abad441962781e0cd83.webp)

Coding agents never visit your website

They read node_modules.
What they remember about your API is from training. Months stale.

Problem 05

## Full transcript

[0:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=1s) [music]

[0:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=12s) >> The talk title, we'll see if it lines up

[0:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=15s) by the end of it, but when we put this

[0:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=17s) talk title in

[0:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=19s) just to be honest with you, so much

[0:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=21s) changes in like 3 days at this point.

[0:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=24s) We'll see how it goes. So, yeah, the

[0:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=27s) whole point of it was that

[0:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=30s) how I got LLMs to understand my open

[0:32](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=32s) source library and what I did to do it

[0:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=36s) well. Is it some kind of scientific

[0:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=38s) background? Am I from a lab? No.

[0:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=43s) That's

[0:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=44s) My slidey clicky thing's not working.

[0:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=46s) So, I just like to say again, I'm just

[0:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=48s) like you. I'm just this side of the

[0:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=51s) stage. I've just hacking hacking it

[0:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=54s) together, figuring out what is useful,

[0:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=57s) what is token efficient, these kind of

[0:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=59s) things. And again, I am British. Please

[1:02](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=62s) don't think my accent makes me an

[1:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=64s) expert.

[1:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=67s) So, for quick contacts, I'm Christopher

[1:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=69s) Burns. I'm the founder of Inth. I

[1:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=72s) created a open source cookie banner

[1:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=74s) library called C15T. That really

[1:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=77s) annoying thing on the internet. That is

[1:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=79s) me.

[1:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=80s) Uh I spoke at Next Conf uh after it

[1:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=83s) started taking off and it had 1.2

[1:26](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=86s) thousand downloads at the time. Now it's

[1:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=87s) closer to 2 million.

[1:29](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=89s) In terms of like statistics, so we just

[1:32](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=92s) checked that, you know, this is not

[1:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=93s) theoretical. This is actual

[1:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=97s) uh something that is succeeding.

[1:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=99s) We have 3 million NPM downloads.

[1:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=101s) 4.5 uh 4.5 45% month-on-month growth.

[1:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=106s) 2.8 thousand websites using it in

[1:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=109s) production from Minify to Z to

[1:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=112s) Inphysical.

[1:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=114s) And

[1:55](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=115s) the whole concept of this talk was

[1:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=118s) uh it goes back to we were doing all

[2:00](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=120s) these things to make our library more

[2:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=123s) efficient. You know, we were batting

[2:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=125s) upwards

[2:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=127s) compared to every other tool. Every

[2:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=129s) other tool was built for marketers and

[2:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=131s) lawyers. We were built for the

[2:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=132s) developer. So, we had to make sure we

[2:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=135s) had a very good developer experience.

[2:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=137s) And we had an onboarding format said,

[2:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=140s) "How did you hear about this?" And we

[2:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=142s) started to get spikes that from April

[2:25](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=145s) 13th, you know, now it is our number one

[2:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=150s) source of inbound is

[2:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=153s) Claude, ChatGPT, Codex, that is ChatGPT,

[2:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=157s) Gemini recommending us.

[2:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=159s) And I like to think of this as, you

[2:42](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=162s) know, the iceberg. You know, we start

[2:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=164s) with the top of C15T

[2:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=166s) and

[2:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=168s) there's many many tools that go into it

[2:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=171s) from, you know, LLMs.txt to site maps to

[2:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=174s) RSS feeds to robot.txts.

[2:56](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=176s) So many micro optimizations that you can

[2:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=178s) do from old methods of running the

[3:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=181s) internet to new methods.

[3:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=183s) And uh

[3:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=185s) how many of you have, you know,

[3:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=189s) made these kind of tools? How many of

[3:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=191s) you have, really put simply, said, "Hey,

[3:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=196s) agents, we need to this to be done." And

[3:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=201s) yeah, it said, "We should install this

[3:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=203s) library." And you've gone, "Okay." Raise

[3:26](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=206s) your hands. How many people have done

[3:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=207s) this?

[3:29](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=209s) Pretty much most people.

[3:32](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=212s) That's a lot of hands.

[3:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=214s) So, what's really funny is that we went

[3:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=217s) from wizards installing our software to

[3:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=219s) agents installing them. And I just went

[3:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=223s) through Y Combinator.

[3:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=225s) And what's really interesting is it if

[3:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=228s) you know who these two people are. These

[3:50](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=230s) are the co-founders of Stripe, the

[3:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=232s) Collison brothers, and they had a really

[3:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=234s) classic saying of like a Collison

[3:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=237s) brothers install.

[3:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=239s) And they would hand you their laptop,

[4:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=241s) and they would install Stripe. These

[4:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=243s) days, it's kind of like just a prompt.

[4:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=246s) Being in Y Combinator, we just give

[4:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=249s) people a prompt. And really what that

[4:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=251s) means is that our very good developer

[4:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=254s) experience primitives are now hitting

[4:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=257s) agent primitives.

[4:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=260s) So,

[4:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=261s) as we was pulling all these things

[4:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=263s) together, there is no one tool that

[4:26](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=266s) fixes everything. I like to think about

[4:29](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=269s) these problems like, you know, Batman's

[4:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=271s) utility belt. Loads of really small

[4:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=273s) things targeted in different areas to

[4:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=276s) get it done.

[4:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=278s) And we built all of these things into

[4:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=281s) C15T because we wanted C15T to be the

[4:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=284s) best developer framework in this tool.

[4:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=287s) Think of it like Stripe Docs. And as we

[4:50](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=290s) was building more and more tools, more

[4:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=292s) and more documentation websites, we

[4:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=294s) actually started abstracting these tools

[4:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=297s) into a side quest

[4:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=299s) that we call Lead Type. So, all of the

[5:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=303s) things that we're going to talk about

[5:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=304s) now are things that we have already

[5:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=307s) solved with this open source framework.

[5:10](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=310s) We have our friends at other developer

[5:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=312s) companies implementing it and seeing

[5:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=314s) similar results about how to like

[5:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=316s) optimize for the agent experience.

[5:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=319s) So, again, this isn't a magic SEO tool.

[5:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=323s) It's actually

[5:25](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=325s) a very non-sexy title, but a

[5:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=327s) framework-neutral docs pipeline.

[5:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=330s) Complex.

[5:32](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=332s) But really, all it basically does is

[5:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=333s) take your .mdx files, you run Lead Type

[5:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=337s) generate, and it will spit out

[5:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=338s) everything for

[5:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=341s) um optimized agent experience for your

[5:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=343s) websites.

[5:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=345s) And the rest of this talk is going to

[5:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=347s) look a bit like a BuzzFeed list, to put

[5:50](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=350s) simply,

[5:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=351s) of these problems, because again, not

[5:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=353s) everybody knows even how to put an

[5:56](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=356s) LLM.txt on their website. So, you know,

[5:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=359s) that comes to the first problem of if

[6:02](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=362s) your docs have hundreds of pages,

[6:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=365s) and how can it navigate them to find the

[6:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=367s) right questions?

[6:10](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=370s) The first solution is obviously an

[6:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=372s) LLM.txt.

[6:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=374s) What we found in our research is that

[6:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=377s) it's much better not to just generate

[6:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=379s) this. It is much better to write your

[6:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=382s) LLM.txt from hand. Obviously,

[6:25](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=385s) AutoRaptor, but write it as you are

[6:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=388s) trying to get the answers across to the

[6:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=391s) LLMs.

[6:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=393s) For about 40 good lines beats 1,000

[6:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=396s) lines of noise from our testing.

[6:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=399s) And that comes to the second issue of

[6:42](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=402s) agents don't know how to browse. They

[6:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=404s) know how to fetch.

[6:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=406s) So, you then need the second part of the

[6:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=409s) solution of the LLMs full.

[6:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=411s) Again, think of this as a sitemap, where

[6:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=414s) it takes the actual page and the links

[6:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=418s) and a short description of what each

[7:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=421s) page is for the LLMs to reference.

[7:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=423s) Again, most people have heard these two

[7:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=425s) solutions.

[7:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=427s) But, where things are starting to get

[7:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=429s) very complicated and we're seeing a lot

[7:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=431s) of optimizations right now, is that HTML

[7:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=434s) is expensive, and why can't we just ship

[7:18](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=438s) markdown to the agents?

[7:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=440s) And we can. And you've seen that

[7:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=444s) everybody has started creating twin MDs.

[7:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=448s) So, that's taking the normal website,

[7:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=450s) such as Next.js quick start, and then

[7:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=453s) having a dot MD on the end of it.

[7:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=457s) And when you load that,

[7:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=460s) it goes to the markdown

[7:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=463s) But, what's really important here, and

[7:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=465s) it's really worth noting, is this line

[7:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=467s) at the bottom. If you look at all the

[7:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=469s) best documentation websites, Minify,

[7:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=473s) Vercel, C15T, pat myself on the back. Um

[7:56](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=476s) they all have this in the header. This

[7:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=479s) is saying to the agents whenever they

[8:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=481s) visit the website that there is an

[8:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=484s) alternative

[8:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=485s) version of this in mark markdown. Again,

[8:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=489s) who actually supports it? Don't ask me.

[8:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=492s) Perplexity, some of the agents, it's all

[8:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=495s) up in the air.

[8:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=497s) And then, the second thing as well

[8:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=500s) is that taking the .mds,

[8:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=504s) you need to make sure that they're

[8:25](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=505s) available through multiple methods. So,

[8:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=508s) one of them is like the .md, so as you

[8:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=511s) like copy it to an agent, you say .md.

[8:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=514s) Another one is just taking the normal

[8:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=518s) um link and then adding a uh redirect

[8:42](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=522s) into your like your Next.js config, so

[8:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=524s) that if it detects an agent has the

[8:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=527s) header of accepting markdown, instead of

[8:50](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=530s) returning the HTML, it will return the

[8:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=533s) markdown. And then, the third one is

[8:56](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=536s) that not all agents can append header

[8:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=539s) tags. So, there's also a URL query of

[9:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=543s) mode equals agent.

[9:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=546s) So, they're the ones that pretty much

[9:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=548s) everybody knows. Um and it's pretty

[9:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=552s) basic internet knowledge at this point.

[9:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=554s) Um

[9:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=555s) but one of the really interesting ones

[9:18](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=558s) is where we're going next. And our

[9:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=560s) tooling is also helping this is that an

[9:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=563s) agent can't ask your website anything.

[9:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=568s) So, we need to think about the web MCP.

[9:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=571s) And this is still very early, but our

[9:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=574s) tool is already uh

[9:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=576s) um exposing three different tools to

[9:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=580s) WebMCP. Search docs, get pages, and ask

[9:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=584s) docs. Again, um our library lead type is

[9:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=588s) pulling all of that context together so

[9:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=591s) an agent can easily ask it the right

[9:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=592s) questions. I think we'll even see a

[9:55](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=595s) future where communication happens over

[9:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=598s) email and there's companies in San

[10:00](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=600s) Francisco building that today.

[10:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=604s) But this is actually the most

[10:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=606s) interesting one and I think the most

[10:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=609s) important one that anybody who has any

[10:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=612s) type of developer module surface, NPM

[10:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=615s) modules, cargo, Python, whatever.

[10:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=619s) Is that the uncomfortable truth is that

[10:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=622s) coding agents are actually never

[10:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=624s) visiting the website if you have a

[10:26](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=626s) library.

[10:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=627s) They're actually visiting the node

[10:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=630s) modules. They read the repo and they

[10:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=633s) read the node modules. They They have

[10:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=636s) previous stale training data

[10:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=639s) and they're trying to work it out on

[10:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=641s) what it can do from the the compiled

[10:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=644s) source.

[10:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=646s) So, again, following what people like

[10:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=649s) Vercel are doing and people who are

[10:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=651s) thought leaders in this industry is that

[10:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=654s) we take the bundled markdown documents

[10:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=659s) and then we also put them in the node

[11:02](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=662s) modules

[11:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=663s) with an agents.md

[11:06](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=666s) file.

[11:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=667s) And the agents.md file basically says,

[11:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=669s) "If you've got a problem, if you've got

[11:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=671s) a question,

[11:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=672s) all the documents are here. Grab them."

[11:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=676s) And we actually see that this has

[11:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=679s) surprisingly real effects. We can see

[11:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=682s) that between many different models

[11:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=684s) almost 50% token saving on instead of

[11:28](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=688s) trying to search the web, find the right

[11:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=691s) tools, pulling the markdown files from

[11:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=694s) your code base. So, if you have a

[11:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=696s) library that's forever changing,

[11:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=700s) then having the node modules built in is

[11:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=703s) a very effective solution. This is also

[11:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=705s) working without any skills, but if you

[11:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=708s) want as well, you can add skills to it

[11:50](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=710s) to say, "Look at the node modules and go

[11:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=712s) from that."

[11:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=714s) And again, just uh doubling down into

[11:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=717s) this point, looking at like the

[11:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=719s) agents.ai and DFile, you can say like

[12:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=721s) when working with Z15 T Next.js library,

[12:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=724s) read the bundles and

[12:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=727s) verify that they match and go from

[12:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=729s) there.

[12:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=731s) So,

[12:12](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=732s) that's really like how we've done it. I

[12:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=736s) don't want to say this is like

[12:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=737s) prescriptive, that I know the answers.

[12:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=741s) If you have documentation websites, or

[12:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=743s) if you have any type of markdown, if

[12:25](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=745s) you're running your own blog, you know,

[12:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=747s) I've been using our package as well on

[12:29](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=749s) our marketing website. Every part of our

[12:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=751s) marketing website also has a markdown

[12:34](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=754s) file. It can be something that's used

[12:36](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=756s) for many things. We're currently just um

[12:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=760s) most people are just using it for

[12:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=761s) documentation.

[12:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=763s) But, you can literally run it and it

[12:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=764s) will pull out all of these extra files.

[12:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=768s) And

[12:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=769s) one of the big things was when I put

[12:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=771s) this talk together,

[12:52](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=772s) you know, we were seeing the results

[12:54](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=774s) that Claude was recommending, but there

[12:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=777s) was not really any like test suites yet,

[13:00](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=780s) or test harnesses, on like is your site

[13:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=783s) agent ready? And Cloudflare brought one

[13:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=785s) of them out.

[13:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=787s) But, my favorite

[13:10](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=790s) is actually one called Aura AI. Um this

[13:13](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=793s) is brand new and it tests a lot.

[13:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=797s) I'm happy to show off score of 59

[13:20](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=800s) because it's constantly changing. 3

[13:22](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=802s) weeks ago, it was a lot higher. And

[13:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=804s) again, this is a forever changing area.

[13:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=807s) So, aura.ai, put in your website, and it

[13:30](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=810s) will start giving you recommendations.

[13:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=813s) It's forever changing. Again, we can

[13:35](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=815s) just stay on top of it. And yeah, this

[13:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=819s) is like one of my final slides is that

[13:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=821s) the slide the

[13:42](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=822s) the market agents, LLMs, everything is

[13:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=826s) forever changing. There is no such thing

[13:48](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=828s) as perfection. When I started making

[13:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=831s) these slides, I got so caught up of

[13:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=833s) like, everyone expects me to be the

[13:55](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=835s) expert here, but I've just been hacking

[13:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=837s) on this problem a little more than you

[13:58](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=838s) guys have so far. So, never get caught

[14:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=841s) with being perfect. Every small little

[14:04](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=844s) increase really does matter. Every small

[14:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=847s) little thing you add really does matter.

[14:11](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=851s) Thank you so much. You can find me on X,

[14:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=854s) Burn Chris, and LinkedIn, and

[14:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=855s) everywhere.

[14:16](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=856s) >> [applause]

[14:17](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=857s) >> WOO!

[14:19](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=859s) I THINK

[14:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=861s) I THINK WE HAVE TIME for one or two

[14:23](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=863s) questions. Yeah, of course.

[14:32](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=872s) >> So, if you were building um

[14:35](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=875s) uh we're a website agency. We work with

[14:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=877s) a lot of startups building like their

[14:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=878s) own websites.

[14:39](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=879s) >> Mhm.

[14:40](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=880s) >> If you were just building a website, not

[14:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=881s) necessarily like developer tool, but

[14:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=883s) just a website to be found, which of

[14:45](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=885s) these methods like would you concentrate

[14:47](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=887s) on if you're starting from scratch?

[14:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=889s) >> Yeah, I think the most important ones,

[14:53](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=893s) and we're starting to see this more and

[14:55](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=895s) more, is

[14:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=897s) trying to provide a dot md file for

[14:59](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=899s) every single page. A lot of CMSs are not

[15:02](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=902s) built in this way. Um and we see this

[15:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=905s) optimization happening more and more

[15:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=907s) where I didn't put in the slide, but

[15:10](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=910s) we're seeing more and more websites

[15:13](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=913s) being visited by agents instead of real

[15:15](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=915s) humans. So, in terms of even like trying

[15:18](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=918s) to be proactive and token efficient, you

[15:21](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=921s) should provide a markdown file if you

[15:24](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=924s) can. Again, a lot of CMS's are not built

[15:27](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=927s) this way. I actually built my own CMS.

[15:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=931s) My name is Chris and I built ChrisCMS,

[15:33](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=933s) short for Christmas. It's a whole It's a

[15:35](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=935s) whole thing my team wishes I never

[15:37](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=937s) built.

[15:38](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=938s) But, it does work and it does bring this

[15:41](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=941s) like token efficiency up. So, yeah, I

[15:44](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=944s) would say llms.txt

[15:46](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=946s) is your first shout.

[15:49](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=949s) llms.txt full form

[15:51](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=951s) .txt. Second, if you if you can, just do

[15:55](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=955s) them manually. Say you're not even

[15:57](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=957s) working on systems that have markdown, I

[16:01](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=961s) still recommend them.

[16:03](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=963s) But, you can always get creative with

[16:05](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=965s) creating these

[16:08](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=968s) files on the on on the go.
