SuperAPI AI Directory

docs

General & other

View repository

Updated 10-15 22:34

About this skill

Docs

Keep the repo legible to agents and humans.

Principles

  • Docs rot silently — every code change is a possible doc change
  • Optimize for decision-relevant, repo-owned truth per token
  • Describe current state; keep history only in migration notes, changelogs, and decisions
  • Lead with supported capabilities and next actions
  • Prefer short, task-shaped sections and structured facts over narrative prose
  • Keep routing docs short and point to deeper docs instead of duplicating them
  • Use repo-relative links for in-repo docs
  • Keep repo docs, agent guidance, and work tracking linked but distinct

Negative-state rule

When rewriting current-state docs, delete every absent, removed, or unprovisioned item, including specific names and category paraphrases, unless it changes a plausible current action. Prior curiosity, old tickets, and "agents ask about it" do not make an absence operational. If a limitation must stay, state one precise boundary and the supported path.

Boundaries

Not docs work: tactical work planning; epic or tracker-ticket creation; boot/readiness setup; baseline PR, issue, contributor, or security policy templates; independent code review; runtime verification.

Workflow

1. Audit the doc surface

Check the files agents and humans actually rely on:

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

Flag stale commands, dead paths, duplicate guidance, routing failures, narrative history, exhaustive negative inventories, and repo-internal details leaking into reader-facing docs.

Before editing:

  • classify the target as repo docs, agent guidance, or a durable spec or decision
  • identify the decision or task each section supports
  • preserve commands, paths, invariants, boundaries, and recovery steps

Use references/source-boundaries.md before writing cross-repo, private workspace, or local-machine facts into checked-in docs.

2. Update routing docs

Keep top-level docs terse and navigational.

  • AGENTS.md should be a compact operating contract and map, not a wiki
  • If the repo uses AGENTS.md, make CLAUDE.md a symlink or @AGENTS.md import instead of maintaining a second authored file
  • README.md should lead with value, quick use, and links to deeper docs
  • Refresh CONTRIBUTING.md and SECURITY.md when they already exist or when moving existing policy out of an overloaded README.md; do not invent baseline policy from scratch
  • For workspace repos, keep one canonical setup doc and let README.md point to it
  • Use the concrete top-level split and section order in references/documentation.md
  • Use reader-facing labels in routing lists; use raw filenames only when the filename matters

3. Update deep docs and specs

Refresh the detailed documents that carry the knowledge.

  • architecture and API docs
  • task guides and runbooks
  • durable feature specs and decision records
  • readiness infrastructure docs after boot, smoke, observability, or isolation changes

Write each updated section as the reader's current source of truth.

For agent-facing or internal docs, follow the selection, structure, and progressive-disclosure guidance in references/agent-first.md.

When the user asks to save a durable rule, prompt, specification, or decision, choose its durable home.

Tactical implementation plans, epics, tracker tickets, and session handoffs are a separate work-tracking operation. When requested alongside docs cleanup, keep the surfaces linked and report the planning work as a distinct next action instead of inventing a checked-in plan home.

For durable feature contracts and decisions, use the selection rules and templates in references/specifications.md.

4. Clean up drift

  • deduplicate repeated facts
  • delete or archive stale docs
  • fix cross-links and moved paths
  • keep naming, labels, casing, commands, and section order consistent
  • keep one canonical home for setup or install commands and replace copied command blocks with pointers
  • prefer reader-facing link text over raw paths unless the path is the point
  • apply the agent-first cleanup filters

5. Validate reality

Verify prose against the repo.

Concrete checks:

  • rg -n "old/path|stale-command" AGENTS.md CLAUDE.md README.md docs/ when paths or commands moved
  • rg -n "<new command|new path|decision keyword>" AGENTS.md CLAUDE.md README.md docs/ to find duplicate or conflicting homes
  • 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
  • for claims sourced from outside the repo, cite or verify the upstream source before making the claim durable
  • compare the before and after facts; every current command, invariant, boundary, and recovery path must remain represented
  • apply the negative-state rule above to the final rewrite

Output

After docs work, report a compact docs footer:

  • files updated
  • verified: command names or path checks, not output logs
  • removed or rewritten: only if stale or duplicated docs changed
  • gaps: remaining doc gaps, or none
  • next: work planning, readiness setup, independent review, runtime verification, or none

Keep the footer to 5 labeled lines or fewer. List changed files once.

References