Start with a durable source
Agent-ready documentation does not begin with another generated corpus. It begins with accurate, well-structured content, stable addresses, explicit permissions, and a publication process people can review.
Use each representation for its job
Semantic HTML serves readers and web discovery. Clean Markdown reduces presentation noise for retrieval. An llms.txt file can map important resources. A content API supports controlled queries, while MCP lets a compatible tool discover and call scoped operations.
- Do not treat llms.txt as an access-control boundary.
- Apply the same visibility rules to HTML, APIs, and MCP.
- Return canonical identifiers and update timestamps.
- Make tool actions read-only unless mutation is explicitly intended and reviewed.
Evaluate answers, not format presence
Test whether representative questions retrieve the right source, preserve important qualifications, and cite a current answer. Shipping an endpoint is implementation; reliable answers are the outcome.