docs（文档）

Docs

Keep the repo legible to humans and agents.

Principles

• Docs rot silently — every code change is a possible doc change
• Documentation is part of the interface; optimize for scanability, rhythm, and visual clarity, not just correctness
• Routing docs stay short; depth lives in docs/
• No duplication when a pointer will do
• Use repo-relative links for in-repo docs; external links are fine in sources and references
• Doc drift is a real failure, not polish debt

Handoffs

• Missing boot, smoke, e2e, logs, or agent-readiness infrastructure → use agent-readiness
• Need to judge existing code, a diff, branch, or PR with evidence → use review
• Need to validate your own completed change on the real surface → use verify

Workflow

1. Audit the doc surface

Check the files humans and agents actually rely on:

• AGENTS.md
• CLAUDE.md
• README.md
• CONTRIBUTING.md
• SECURITY.md
• docs/
• plans, specs, runbooks, decision docs

Flag stale commands, dead paths, duplicated guidance, routing failures, and places where filenames or implementation order are leaking into the visible docs surface.

2. Update routing docs

Keep top-level docs terse and navigational.

• AGENTS.md should be a table of contents, not a wiki
• If the repo uses AGENTS.md, keep CLAUDE.md at the same level as a symlink to AGENTS.md instead of maintaining a second authored file
• README.md should lead with value and the fastest path to use the project
• CONTRIBUTING.md should hold contributor setup, validation, and workflow
• SECURITY.md should hold private-first vulnerability reporting guidance
• Push detail downward instead of bloating top-level files
• For coordination or workspace repos, keep one canonical setup doc and let README.md point to it instead of repeating the full bootstrap flow inline
• Use the concrete top-level split and section order in references/documentation.md instead of inventing a new shape every time
• Keep visible docs copy human-facing and task-ordered; let the reference file own the detailed labeling and scannability rules
• Prefer terse routing over narrative sprawl, for example README.md should link to deeper docs instead of re-explaining them inline

3. Update deep docs and specs

Refresh the detailed documents that actually carry the knowledge.

• architecture and API docs
• task guides and runbooks
• feature specs, plans, and decision records
• readiness infrastructure docs after agent-readiness changes

For new features, use the directory layout and templates in references/structuring.md — specs, plans, and decisions each have their own shape.

4. Clean up drift

• deduplicate repeated facts
• delete or archive stale docs
• fix cross-links and moved paths
• keep naming and commands consistent across files
• keep one canonical home for setup or install commands in workspace-style repos, and replace copied command blocks elsewhere with short pointers
• normalize visible labels, casing, and section order when the docs read like a file tree instead of a user guide

Example — fixing a stale path after a rename:

 # AGENTS.md
-- Run `scripts/bootstrap.sh` to set up the dev environment.
+- Run `scripts/setup.sh` to set up the dev environment.

5. Validate reality

Do not trust prose. Check that commands, file paths, and entry points still match the repo.

Concrete checks:

• rg -n "old/path|stale-command" AGENTS.md CLAUDE.md README.md docs/ when paths or commands moved
• test -e <path-from-docs> before keeping a file reference
• test! -e AGENTS.md || {test -L CLAUDE.md && test "$(readlink CLAUDE.md)" = "AGENTS.md";} when normalizing agent entrypoints

Output

After docs work, report:

• files updated
• stale or duplicated docs removed or rewritten
• commands or paths verified
• remaining doc gaps
• any handoff needed to agent-readiness, review, or verify

References

• references/documentation.md — AGENTS.md shape, scoped rules, README patterns, doc hygiene
• references/specifications.md — feature specs, conformance tests, spec drift, SDD trade-offs
• references/structuring.md — directory layout, templates, and naming for specs, plans, and decisions