Making Documentation AI-Agent Ready with llms.txt
RevoGrid's team explains how llms.txt, llms-full.txt and Skill bundles work together to help AI coding agents find accurate documentation.
AI coding agents such as Codex, Cursor, Claude Code and Copilot are now routinely asked to install packages, wire up APIs, or migrate code. Since documentation has traditionally been built for human readers, an agent starting from a bare URL often can't find the right entry point and ends up guessing. The RevoGrid team addresses this by adopting llms.txt, the lightweight discovery format proposed by Jeremy Howard in 2024, which points agents to installation guides, API references, examples and migration notes.
The piece is careful to note that llms.txt is not a magic AI-SEO trick. Despite adoption by OpenAI, Anthropic, Cloudflare, Stripe and Vercel, there's no evidence it improves rankings or model accuracy on its own — some studies even found published files receive very few crawler requests. Still, the team argues it's worth publishing when it can be generated cheaply from canonical docs and treated as one piece of a broader agent-documentation strategy.
Because different agents behave differently, RevoGrid layers several formats: llms.txt for discovery, llms-full.txt as a plain-text fallback, an MCP server for live retrieval, and a split Skill bundle with one reference file per doc page for large documentation sets. This Skill format follows Anthropic's open Agent Skills standard and OpenAI's Codex SKILL.md convention, both built around progressive disclosure — giving agents a short description first and full detail only when needed.
For engineers, the takeaway is that documentation now needs to serve agents as a first-class audience. Clear, up-to-date, well-scoped docs (marking what's core versus Pro/Enterprise) directly reduce the odds of an agent generating broken or stale code.