# How We Got LLMs to Recommend Our Open Source Library Christopher Burns, Inth | AI Engineer World's Fair 2026 | 16:27 Source: https://www.youtube.com/watch?v=V_5bn4q-vAI Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/how-we-got-llms-to-recommend-our-open-source-library Published: 2026-08-26 Tags: agents, coding-agents, developer-experience, mcp ## TL;DR - Christopher Burns traced c15t's growing inbound to Claude, ChatGPT, Codex, and Gemini recommending the library to developers. - A hand-written llms.txt, linked markdown pages, and several ways to request markdown give agents smaller, clearer documentation to fetch. - Coding agents often read repositories and node_modules instead of visiting documentation sites, so shipping bundled markdown and AGENTS.md with a package can save close to half the tokens. ## Summary Christopher Burns explains how Inth made its open source consent banner library, c15t, easier for language models and coding agents to understand. The work started with good developer experience and expanded into many small documentation changes. Burns describes a framework-neutral pipeline that generates agent-facing files from MDX, while recommending that llms.txt itself be written by hand. He also explains why documentation should be available as markdown through a file suffix, content negotiation, and a query parameter. For libraries, the most important destination may be the package itself. Coding agents often inspect a repository and node_modules rather than opening a documentation site, so Inth ships bundled markdown and an AGENTS.md file in the package. Burns says this produced close to 50% token savings in testing. He is candid that the methods keep changing and that no site remains permanently agent-ready. ## Key ideas ### LLMs became c15t's largest source of inbound [02:00](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=120s) Inth added an onboarding question asking how people heard about c15t. From April 13 onward, the answers began showing a new pattern. Claude, ChatGPT, Codex, and Gemini were recommending the open source cookie consent library, making LLM referrals its largest source of inbound. Burns connects this to the library's focus on developer experience rather than tools built mainly for marketers and lawyers. He says the results were visible in actual usage, with c15t approaching 2 million downloads and 2.8 thousand production websites at the time of the talk. ### Agent discoverability comes from many small changes [02:31](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=151s) Burns does not attribute the referrals to one search optimization trick. He compares the work to Batman's utility belt, with small tools aimed at different parts of the agent experience. The list includes llms.txt, sitemaps, RSS feeds, and robots.txt, along with other documentation changes. Inth eventually abstracted these techniques into Lead Type, a framework-neutral documentation pipeline. It takes MDX files and generates files intended to make a documentation site easier for agents to use. ### A hand-written llms.txt gives agents a compact route through the docs [06:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=367s) For documentation sites with hundreds of pages, Burns recommends starting with llms.txt. His testing found that writing it by hand works better than generating it automatically. He says roughly 40 good lines beat 1,000 lines of noise because the file needs to communicate useful answers and direct an agent toward the right pages. A second file, llms-full.txt, works more like a sitemap. It lists the actual pages, their links, and a short description of what each page is for. ### Markdown should be available through several request patterns [07:07](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=427s) Burns argues that HTML is expensive for agents and recommends serving a markdown version of every documentation page. One approach adds .md to the page URL. A second uses content negotiation, returning markdown when the request accepts it. A third adds a query parameter such as mode=agent for agents that cannot set request headers. Documentation sites can also announce the markdown alternative in the page header. Burns says support varies across agents, so offering multiple access methods is practical. ### Web MCP lets an agent search and ask the documentation directly [09:14](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=554s) Burns describes Web MCP as an early way for an agent to interact with a documentation site instead of only fetching pages. Inth's tooling exposes three tools: search docs, get pages, and ask docs. Lead Type gathers the documentation context behind those tools so an agent can ask targeted questions. Burns presents this as an emerging part of the system rather than a finished standard. He also mentions that some companies are exploring communication between agents and companies over email. ### Library documentation needs to ship inside the package [10:09](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=609s) Burns calls it an uncomfortable fact that coding agents with a library often never visit its documentation website. They inspect the repository, node_modules, and compiled source, while also relying on stale training data. Inth responds by putting bundled markdown documentation into node_modules and adding an AGENTS.md file that tells the agent where the documents are. Burns says this approach produced close to 50% token savings across several models compared with searching the web and retrieving the right material. ### Agent readiness needs repeated testing because the target keeps moving [12:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=763s) Burns says that when he prepared the talk, there were few test suites for checking whether a site was ready for agents. He mentions a Cloudflare tool and recommends Aura AI, which gives a site recommendations after testing it. His own score was 59, and he says it had been much higher three weeks earlier. The score illustrates his broader point: agent behavior and evaluation tools change continually. He advises making incremental improvements rather than treating an implementation as finished. ### For an ordinary website, start with markdown pages and llms.txt [14:43](https://www.youtube.com/watch?v=V_5bn4q-vAI&t=883s) In the question period, a website agency asks what to prioritize for a normal startup website rather than a developer library. Burns recommends providing a .md file for every page where possible, since agents are increasingly visiting websites and markdown is more token-efficient. He acknowledges that many CMSs do not support this directly. His fallback is to create the files manually or generate them as needed. He places llms.txt near the start of the process, followed by llms-full.txt, and again recommends writing the important files by hand. ## Notable quotes - "Our number one source of inbound is Claude, ChatGPT, Codex, that is ChatGPT, Gemini recommending us." (02:20) - "For about 40 good lines beats 1,000 lines of noise from our testing." (06:06) - "Coding agents are actually never visiting the website if you have a library." (10:09) - "We can see that between many different models almost 50% token saving." (11:03) - "There is no such thing as perfection." (13:33) ## Tools & references mentioned - Inth - c15t - Lead Type - Claude - ChatGPT - Codex - Gemini - Stripe - Y Combinator - Vercel - Next.js - Perplexity - Web MCP - Cloudflare - Aura AI - NPM - node_modules - AGENTS.md - llms.txt - llms-full.txt ## Who should watch - You maintain an open source package and want coding agents to understand its API without relying on old training data. - Your documentation site has many HTML pages and you need agents to find the right page with fewer fetches and tokens. - You build websites for startups and want a practical starting point for making ordinary pages easier for agents to read.