agent-docs AGENTS.md

AGENTS.md

Overview

• agent-docs ships three skills (/gather-agent-docs, /use-agent-docs, /refresh-agent-docs) that give a coding agent a curated, on-disk reference for the project's key topics — answer locally before web-searching.
• Skills are file-based and harness-agnostic: the same skills/ tree works in Claude Code, Gemini CLI, GitHub Copilot, and OpenCode.
• End-user state lives in .agents/docs/INDEX.md plus one <topic>/DOC.md per topic and optional <SUBTOPIC>.md sibling files. Each page opens with YAML frontmatter (last_verified, optional upstream_commit, sources.{docs,repo,changelog,release}); INDEX.md's per-topic date is the oldest last_verified among the topic's files.

Conventions

• Line caps. Skill files exist to be loaded by an agent on every run; verbosity directly costs context budget. Targets:

  • skills/*/SKILL.md — ≤ 110 lines
  • Project-root markdown (README.md, AGENTS.md) — uncapped, but tighten when it sprawls

  When a file approaches its cap, prefer cutting motivational/explanatory prose over removing structural rules. Each rule should be stated once, in the imperative, in the file the agent loads when it needs it. The "why" lives in commit messages and PR descriptions.

• Lint. Run npm run lint to validate changes — Prettier formatting plus markdownlint-cli2. pre-commit and CI run the same hook.

• Voice. Skill bodies address the agent in the imperative ("Read X, then write Y"). Avoid second-person flavor text aimed at the human reader — that belongs in the README.

• Capture vs. link. The gather-agent-docs rules apply to this repo's own behavior: build the documentation map by default (T1: page title + URL + one-line hook); capture paste-ready snippets only on request (T2: capped at ~5 per topic, each with a per-block source link); link out (never capture) for volatile values like model IDs and pricing (T3); copy verbatim from upstream just fetched in this session (never paraphrase, infer, or auto-complete); keep pages reusable across projects.