How We Got LLMs to Recommend Our Open Source Library — Christopher Burns, Inth
For fast-changing libraries, ship concise agent-facing docs inside the package as well as on the web; agents may inspect installed code before visiting your site.
C15T reports **3 million npm downloads**, with LLM recommendations now its largest inbound source. Its approach combines a hand-written llms.txt, Markdown variants of pages, WebMCP tools, and bundled package docs referenced by AGENTS.md.
Start with a concise, manually curated llms.txt, expose page-level Markdown through multiple access paths, and bundle current docs with rapidly changing libraries. The speaker’s testing found roughly **40 useful lines** beat 1,000 noisy ones and package-local docs saved nearly **50% of tokens** across several models.
C15T reports **3 million npm downloads**, with LLM recommendations now its largest inbound source. Its approach combines a hand-written llms.txt, Markdown variants of pages, WebMCP tools, and bundled package docs referenced by AGENTS.md. Start with a concise, manually curated llms.txt, expose page-level Markdown through multiple access paths, and bundle current docs with rapidly changing libraries. The speaker’s testing found roughly **40 useful lines** beat 1,000 noisy ones and package-local docs saved nearly **50% of tokens** across several models. The talk does not isolate which change caused recommendation growth, and support for Markdown headers, URL conventions, WebMCP, and agent-readiness tests remains inconsistent. Treat each mechanism as an experiment and measure it against real agent tasks.
This extends context engineering from repository use to whether agents can discover, understand, and recommend a library at all. It favors concise curated entry points and package-local, current documentation over exhaustive context, while treating llms.txt, Markdown exposure, and WebMCP as separately measurable distribution experiments. The recommendation growth and token savings remain local evidence because attribution and format support are unresolved.