agentleFS
HyperVonHyperVon/agent-guidance-kit/skills/documentation-review/SKILL.md

documentation-review

Skill1 starsChanged 15 days agoApache-2.0View on GitHub

Audit and update project documentation against current implementation, build/configuration, tests, CI, and authoritative external references when needed. Use for documentation audits, documentation sync, stale or incorrect claims, runbooks, guides, READMEs, diagrams, or project-local guidance. Do not use for missing domain depth, anti-patterns, or checklist gaps in a skill (use skill-reviewer); this skill verifies factual agreement with source/build/CI truth.

3 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

12 rules. 0 are common, 10 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 10

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When sources disagree, do not guess.

    line 34
  2. Distinctive:

    Map claims to evidence and audience: Check setup, commands, versions, interfaces, behavior, security/safety statements, diagrams, links, …

    line 44
  3. Distinctive:

    Apply the approved slice: Keep diffs minimal, preserve the document's purpose and tone, link to a canonical explanation instead of …

    line 67
  4. Very specific:

    Verify that every relative file link (e.g., [link text] targeting docs/guide.md) resolves to a tracked file in the repository from the …

    line 78
  5. Very specific:

    Verify that heading anchors (e.g., [link text] targeting #section-name) match the exact slugified header text in the target document.

    line 79
  6. Distinctive:

    When the user has just shipped a feature, refactor, or dependency bump and asks to touch the relevant docs (not a full audit), use this …

    line 84
  7. Distinctive:

    Keep the change minimal: update the one or two owned documents, link to the canonical source instead of duplicating it, and preserve tone …

    line 96
  8. Distinctive:

    Use the repository's size labels when they exist.

    line 103
  9. Distinctive:

    Narrative bloat: Do not add conversational fluff, promotional adjectives, or tutorial walkthroughs to technical reference documents.

    line 117
  10. Distinctive:

    Internal leak: Do not copy internal test fixtures, private developer paths, or internal-only helper functions into public API guides.

    line 118
  11. Distinctive:

    Duplicate truth: Do not copy full policy or procedural text across multiple documents.

    line 119
  12. Distinctive:

    Do not treat a request to review as permission to edit.

    line 121

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Documentation Review

Audit documentation claims against evidence, make the smallest approved correction, and leave an exact verification record. This skill owns document truth and document quality; it does not own product-code fixes, test hardening, architecture decisions, deployment, or remote-system mutations.

Authority and source of truth

Separate authority for the workflow from evidence for factual claims:

  1. The user's request and repository-local rules define scope, permissions, and approval requirements.
  2. Current implementation, public interfaces, build/configuration, tests, generated contracts, CI definitions, and observed behavior are the primary evidence for claims about the repository.
  3. A named primary external reference may establish an external fact; record its source and date when the fact can change.
  4. Existing documentation, changelog history, examples, and screenshots are evidence to check, not authority over current reality.

When sources disagree, do not guess. Classify the mismatch, identify the strongest current evidence, state uncertainty, and either correct the document or defer it. Never broaden a guarantee beyond what the evidence supports.

Workflow

  1. Scope and inventory. Identify each document's intended audience and purpose. Read the requested documents and only the implementation, build/configuration, tests, CI, generated artifacts, or primary references needed to verify their claims. Record the paths read.
  2. Map claims to evidence and audience. Check setup, commands, versions, interfaces, behavior, security/safety statements, diagrams, links, screenshots, and status claims. Classify each finding as WRONG, STALE, MISSING, ORPHAN, UNVERIFIED, or BROKEN. A true statement can still be an ORPHAN when it is placed for the wrong audience or duplicates a canonical owner. Keep a public README focused on understanding, installing, and using the project; move detailed exclusions, trust models, and internal rationale to design or security documentation unless users need them to avoid an immediate misuse.

    Code snippet and flag verification: For every code snippet, configuration sample, or CLI command in documentation:

    • Export & signature match: Verify that imported symbols, class names, method signatures, and parameter names match current source exports exactly.
    • CLI flag validation: Compare flags against CLI help output (--help) or CLI argument parser definitions. Mark flags that no longer exist or changed default values as WRONG or STALE.
    • Config syntax validation: Parse configuration examples (YAML, JSON, TOML) against the current parser schema or configuration models in the repository.
    • Screenshot and diagram verification: For every screenshot or diagram, confirm it reflects the current UI or behavior (run the app/component or inspect the rendered output) rather than an earlier version. A screenshot that shows an outdated layout, old text, or removed feature is STALE; changing its caption is not verification.
  3. Report before broad edits. For each finding, give the document location, claim, evidence anchor, impact, smallest correction, and size. Preserve unrelated work and do not invent new documentation scope.
  4. Apply the approved slice. Keep diffs minimal, preserve the document's purpose and tone, link to a canonical explanation instead of duplicating it, and keep examples generic and secret-free. If a code or test change is needed, hand it to the owning skill rather than hiding it in a doc edit.
  5. Verify the saved result. Re-read changed sections, validate links and syntax, render or validate changed diagrams with the repository's supported tooling, and run the repository's documented documentation checks. A stale visual asset is a separate finding; changing its caption is not visual verification.

    Link and anchor verification:

    • Verify that every relative file link (e.g., [link text] targeting docs/guide.md) resolves to a tracked file in the repository from the source document's directory.
    • Verify that heading anchors (e.g., [link text] targeting #section-name) match the exact slugified header text in the target document.
    • Report any link pointing to an untracked, moved, or deleted file as BROKEN.

Incremental sync after a change set

When the user has just shipped a feature, refactor, or dependency bump and asks to touch the relevant docs (not a full audit), use this incremental path instead of the full workflow above:

ChangeUpdate
Feature, stack, command, or layoutREADME.md (setup, usage, package tree)
Architecture, protocol, or milestone orderArchitecture doc and an ADR (for example under docs/adr/) when the change is architectural
Operation or troubleshooting commandRunbook or troubleshooting guide
User-facing CLI, UI, or operator workflowUser guide and linked runbook sections
Dependency or container versionDocumentation that references the version or pin; use manifests and lockfiles as source-of-truth evidence

Keep the change minimal: update the one or two owned documents, link to the canonical source instead of duplicating it, and preserve tone and audience. For a whole-repository factual audit (missing, wrong, stale), use the full workflow above.

Approval gate

Use the repository's size labels when they exist; otherwise use these defaults:

  • S — One directly evidenced wording, link, or path correction. Apply only inside the user's explicit scope and report it.
  • M — Several documents, a new section, a workflow or compatibility claim, or a broad rewrite. Present the exact files, claims, evidence, and revert shape; wait for explicit approval.
  • High-risk — Security, privacy, safety, data handling, migrations, compatibility guarantees, operational commands, or claims that could cause costly action. Stop for explicit approval, even if the diff is small, and include focused evidence and compensating verification.

Documentation review anti-patterns

  • Narrative bloat: Do not add conversational fluff, promotional adjectives, or tutorial walkthroughs to technical reference documents.
  • Internal leak: Do not copy internal test fixtures, private developer paths, or internal-only helper functions into public API guides.
  • Duplicate truth: Do not copy full policy or procedural text across multiple documents; link to the canonical owner instead.

Do not treat a request to review as permission to edit. Do not create or modify issues, pull requests, releases, shared infrastructure, or other remote state automatically.

Completion contract

Return this report, with no empty or implied fields:

Documentation review
Scope: <requested documents and boundaries>
Sources read: <paths or named authoritative references>
Findings: <count by WRONG/STALE/MISSING/ORPHAN/UNVERIFIED/BROKEN>
Changes: <each changed path and section, or "none">
Verification:
  - <exact check or command> — PASS|FAIL|SKIPPED|BLOCKED; <result/evidence>
  - <link/diagram/render check, if applicable> — PASS|FAIL|SKIPPED|BLOCKED; <result>
Unresolved or deferred: <item, reason, owner/approval needed, or "none">
Changed paths: <exact paths>

Report skipped optional tooling as SKIPPED, not PASS; report unavailable required evidence as BLOCKED. Do not claim completion until every changed document has been re-read and every applicable check is accounted for.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.