Stats
Actions
Tags
From skills
Audit a codebase for documentation that has drifted out of sync with the code — stale vocabulary, violated decisions, README/behavior claims the code no longer backs — and surface each as a finding. Static-first, no tool run required. Use when the user wants to check the docs are still accurate, find stale docs, or asks "is the documentation still accurate?" / "what drifted from the docs?".
How this skill is triggered — by the user, by Claude, or both
Slash command
/skills:audit-docs [path or area to focus on, or leave blank for the whole codebase][path or area to focus on, or leave blank for the whole codebase]The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Find where the **prose no longer matches the code** — documented claims the implementation has outgrown — and surface each as a finding. Run the shared audit method in [../AUDIT-METHOD.md](../AUDIT-METHOD.md) with the lens, sub-dimensions, and dimension below.
Find where the prose no longer matches the code — documented claims the implementation has outgrown — and surface each as a finding. Run the shared audit method in ../AUDIT-METHOD.md with the lens, sub-dimensions, and dimension below.
The aim is factual drift — a doc that contradicts the code — not writing quality. A stale claim that misleads a reader or an agent is the target; an awkward sentence that is still true is not.
The project's documentation is both the claims to check and the vocabulary finding titles use — whatever the in-context project CLAUDE.md points to. With none present, the audit degrades to README and docstring claims; if there is nothing documented to contradict, report nothing rather than invent docs to audit.
Walk the documentation and collect concrete, checkable assertions — a defined term's meaning, a recorded decision, a README's described command or flag, a docstring's stated behavior. Skip aspirational or roadmap prose: only claims the code is supposed to back right now. Rank by how badly a stale claim would mislead.
For each claim, find the code that should make it true. A claim with no corresponding code is itself a signal (dead doc). Trace whether the code actually honors the claim — a term still used with the same meaning, an ADR's decision still implemented, a documented flag still wired, the described output still produced. A doc and code merely sitting near each other is not agreement.
Drop low-confidence noise — a doc that is loose but not wrong is not a finding.
docs-drift (the sub-dimension goes in the title/evidence, e.g. "stale vocabulary term …", "README flag … no longer wired" — never as a new top-level dimension)enhancement ("realign the doc with X"); a doc that actively misleads into wrong behavior is a bugPer ../AUDIT-METHOD.md → Handover.
capture — dedups against open issues, culls, files survivors as needs-triagenpx claudepluginhub krixon/skills --plugin skillsGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.