docs-audit · git:20261007.158dade · 2026-10-07 · sha256 6467c9914d62a3ba
docs-audit git:20261007.158dadeA
Immutable. This exact content never changes and is served at /api/v1/blob/6467c9914d62a3ba.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_3nalymgsipemhdyi/6467c9914d62a3ba95e248b4ab032f66c3e8560b3d3f0035da5eeb20d3d4fb8b.json
Then /plugin install docs-audit@markdownregistry-3nalymgsipemhdyi in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: docs-audit description: Assess an existing Box documentation page (docs/user/, docs/design/, or a crate README.md) for accuracy against the code and its tests, placement, and voice, and recommend what to fix. Use before a rewrite, when a page may have drifted from the code, or when asked to "audit this page", "is this page still true", or "what is wrong with this doc". --- # Documentation audit Assess one existing page, and produce a structured assessment with the fixes in priority order. **How this differs from `docs-reviewer`.** The reviewer is a voice gate for a draft. The audit is for a page that already exists: it checks each claim against the code and its tests, finds the gaps, and decides what needs work. Use the audit to decide *what to fix*. Use the reviewer to decide *that a draft can ship*. ## Inputs - **Target page**: a path in the repository. - **Context** (optional): an issue, a changed behavior, or a concern. ## Process ### Step 1: Read and classify Read the page. Classify its guide and its content type. A page that mixes types is itself a finding, and it is the most common structural problem. ### Step 2: Accuracy Follow [`claim-verification.md`](../../references/claim-verification.md). List each claim on the page, and check it against the code and the test that pins it. **The code is the authority.** Where the page and the code disagree, the page is wrong. Name the symbol or the test that the page disagrees with. Never recommend a code change to agree with a page. Check each of these: - Each `box.toml` key, type, and default against the configuration parser. - Each CLI verb, flag, and exit code against the argument parser. - Each policy action and field against `strands-box policy generate-schema`. - Each binary name against what ships (`Cargo.toml`, and the staging step in `.github/workflows/deploy-box-artifact.yml`). - A test still pins each claim. A test name on a design page is itself a finding. - Each `decisions.md` anchor still exists, and the decision still says what the page says. - Each relative link resolves. - Each claim that a test pins is not stated more broadly than the test. ### Step 3: Voice Walk the five layers of [`voice-guide.md`](../../references/voice-guide.md): structure, framing, register, hard constraints, and authenticity. Run `.agents/skills/docs-reviewer/check-prose.sh <page>` for the mechanical part. ### Step 4: Stands alone Check the list in "Readers that are people and readers that are agents" in the voice guide. ### Step 5: Comparison (optional) Run this step only for a design page about a topic that other sandboxes also document: the architecture, the security model, containment, egress, or credentials. Read one to three pages from comparable projects, for example the gVisor architecture guide, the Firecracker design document, or a coding-agent sandbox's security page. Cite each URL and the date you read it. Give one observation for each of coverage, depth, and structure. Copy structure or coverage only, never voice. ## Output format ```markdown ## Docs audit: <page title> **Path:** <path> **Guide and type:** <guide>, <type> **Verdict:** Strong | Needs work | Significant gaps | Layer | Score | Key finding | |---|---|---| | Structure | ... | ... | | Framing | ... | ... | | Register | ... | ... | | Constraints | ... | <count of hits> | | Authenticity | ... | ... | ### Accuracy issues - <The claim, the line, and the symbol or the test it disagrees with.> Or "None found". ### Stands-alone issues - <Each issue.> Or "Passes". ### Recommended actions 1. <The fix with the highest priority.> 2. ... ### Comparison (if run) - <One observation per axis, with the source.> ``` ## What this skill does not do - Rewrite the page. That is `docs-writer`. - Edit any file, or commit.