{ "video": { "id": "V_5bn4q-vAI", "title": "How We Got LLMs to Recommend Our Open Source Library — Christopher Burns, Inth", "duration": 987, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "V_5bn4q-vAI", "title": "How We Got LLMs to Recommend Our Open Source Library — Christopher Burns, Inth", "one_liner": "The founder of the C15T cookie-banner library explains the concrete, unglamorous stack of micro-optimizations — hand-written llms.txt, markdown twins of every page, Web MCP tools, and bundled docs plus an agents.md shipped inside node_modules — that made Claude, ChatGPT, Codex and Gemini his number one source of inbound.", "summary": "Christopher Burns, founder of Inth and creator of the open source cookie banner library C15T, argues that good developer experience primitives are now colliding with agent primitives: installs used to be a 'Collison brothers install', now they're just a prompt. He walks through a BuzzFeed-style list of fixes — hand-written llms.txt, llms-full.txt as an agent sitemap, .md twins served three ways, Web MCP tools, and most importantly shipping bundled markdown docs plus an agents.md into node_modules because coding agents never visit your website. All of it is packaged in a framework-neutral docs pipeline they open sourced as Lead Type, and he insists nothing here is settled: the area changes constantly and every small increment matters.", "key_points": [ "C15T's traction is real, not theoretical: 3 million NPM downloads, 45% month-on-month growth, 2,800 websites in production (Minify, Z, Inphysical); it was 1.2 thousand downloads when he spoke at Next Conf and is closer to 2 million now.", "Their onboarding form's 'How did you hear about this?' field started spiking on April 13th — LLM recommendations from Claude, ChatGPT, Codex and Gemini are now their number one source of inbound.", "There is no single tool that fixes this — he frames it as Batman's utility belt of small targeted optimizations spanning llms.txt, sitemaps, RSS feeds and robots.txt; the tooling was abstracted out of C15T into an open source, framework-neutral docs pipeline called Lead Type that takes your .mdx files and generates everything on `leadtype generate`.", "Write llms.txt by hand rather than generating it — from their testing, ~40 good lines beats 1,000 lines of noise — then add llms-full.txt as a sitemap-style index of pages, links and short per-page descriptions, because agents don't browse, they fetch.", "HTML is expensive, so ship markdown: publish a .md twin of every page (e.g. the Next.js quick start with .md appended), advertise it with the alternate-markdown line in the page header that Minify, Vercel and C15T all have, and expose it three ways — the .md suffix, a Next.js config redirect triggered by an Accept-markdown header, and a ?mode=agent URL query for agents that can't append headers.", "An agent can't ask your website anything, so C15T's tooling already exposes three Web MCP tools: search docs, get pages, and ask docs; he expects a future where this communication happens over email and says SF companies are building that today.", "The uncomfortable truth for anyone with a module surface (NPM, cargo, Python): coding agents never visit the website — they read the repo, the node_modules and the compiled source, plus stale training data. Following what Vercel and others do, they bundle the markdown docs into node_modules with an agents.md file pointing at them, and measure almost 50% token savings across many models versus searching the web. It works without skills, and skills can be layered on top.", "Agent-readiness test harnesses barely existed when he built the talk; Cloudflare shipped one, but his favourite is Aura AI (aura.ai) — he showed a score of 59 and noted it was higher three weeks earlier because the area keeps changing." ], "takeaways": [ "Hand-write your llms.txt (autoformat, but write it for the answers you want the LLM to give) and pair it with llms-full.txt — those are the first two things to do, even on a plain marketing site or blog.", "Serve a markdown version of every page and make it reachable three ways: .md suffix, Accept-header content negotiation, and a ?mode=agent query param — and declare it via the alternate-markdown header line.", "If you ship a library, put your bundled markdown docs plus an agents.md inside node_modules (or your language's equivalent), telling the agent where the docs are and to verify them against the bundles — that's where coding agents actually look, and it cut token usage nearly in half.", "Run your site through an agent-readiness harness like aura.ai (or Cloudflare's) and treat the score as a moving target rather than something you finish.", "Don't wait for a perfect, settled strategy — every small addition measurably helps, and the landscape changes in days." ], "topics": [ "agents", "llms.txt", "documentation", "developer experience", "mcp", "open source", "seo", "token efficiency" ], "tools": [ "C15T", "Inth", "Lead Type", "NPM", "Next.js", "Claude", "ChatGPT", "Codex", "Gemini", "Perplexity", "Stripe", "Y Combinator", "Vercel", "Minify", "Cloudflare", "Aura AI (aura.ai)", "ChrisCMS", "Web MCP", "agents.md", "llms.txt", "Z", "Inphysical" ], "quotes": [ { "text": "we went from wizards installing our software to agents installing them", "at": "03:37", "url": "https://www.youtube.com/watch?v=V_5bn4q-vAI&t=217s" }, { "text": "For about 40 good lines beats 1,000 lines of noise from our testing.", "at": "06:33", "url": "https://www.youtube.com/watch?v=V_5bn4q-vAI&t=393s" }, { "text": "agents don't know how to browse. They know how to fetch.", "at": "06:42", "url": "https://www.youtube.com/watch?v=V_5bn4q-vAI&t=402s" }, { "text": "the uncomfortable truth is that coding agents are actually never visiting the website if you have a library", "at": "10:19", "url": "https://www.youtube.com/watch?v=V_5bn4q-vAI&t=619s" } ], "words": 2653 }, "summary_url": "/#V_5bn4q-vAI", "transcript": { "html": "/transcripts/V_5bn4q-vAI.html", "txt": "/transcripts/V_5bn4q-vAI.txt", "vtt": "/transcripts/V_5bn4q-vAI.vtt" } }