agentleFS
gtappsgtapps/claude-code-hermit/.agents/skills/docs-drift/SKILL.md

docs-drift

Skill84 starsChanged 5 days agoMITView on GitHub

Documentation-drift audit for this plugin monorepo. Audit either each scoped plugin's unreleased changelog claims or its latest shipped release from the previous reachable plugin-name--v* tag, then verify that plugin READMEs, docs, CLAUDE.md or AGENTS.md files, and root documentation tell the truth for the matching snapshot. Report meaningful contradictions with concrete proposed edits and wait for approval before changing files. Use when preparing or verifying a release, when asked to check the latest version, or when the user says "docs drift", "check the docs", "are the docs up to date", "docs audit", "did the docs keep up", or whether documentation matches released or unreleased changes. Complements pre-release-review, which checks changelog claims against code; this skill checks changelog claims against docs.

6 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

23 rules. 0 are common, 18 are distinctive, and 5 are very specific to this project.

  • 0

    Common

    Most of the 5 edited versions we read keep these.

  • 18

    Distinctive

    This author's own choices. Worth a look.

  • 5

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Verify that documentation tells the truth about either the changes about to ship or the latest version already shipped.

    line 8
  2. Distinctive:

    Keep the audit read-only. Do not edit, commit, tag, push, publish, or release until the user explicitly approves specific findings.

    line 11
  3. Distinctive:

    Use unreleased mode by default.

    line 24
  4. Distinctive:

    Never mix released and unreleased windows in one claim set.

    line 33
  5. Distinctive:

    If the user named a plugin, limit scope to that slug after verifying it exists and has unreleased content.

    line 47
  6. Distinctive:

    Use the working tree as the documentation snapshot.

    line 52
  7. Very specific:

    Start from all plugins/*/CHANGELOG.md slugs, or only the user-named slug.

    line 57
  8. Very specific:

    Use only reachable <slug>--v* tags.

    line 77
  9. Distinctive:

    Use that diff only in latest-release mode and only for claim discovery.

    line 119
  10. Distinctive:

    Do not manufacture a finding for every bullet.

    line 136
  11. Very specific:

    Unreleased: use rg -n against the working tree.

    line 158
  12. Very specific:

    Latest release: use git grep -n <pattern> "$latest" -- <paths> and git show "$latest:<path>".

    line 160
  13. Distinctive:

    Never silently substitute current documentation for a released snapshot.

    line 165
  14. Distinctive:

    When more than two plugins are in scope and agent delegation is available, delegate one plugin audit per worker.

    line 167
  15. Distinctive:

    Keep the root-file sweep in the main session because it needs cross-plugin deduplication.

    line 173
  16. Distinctive:

    If it is still stale, propose an edit to the current file.

    line 180
  17. Very specific:

    If it is already corrected, report it under Fixed after release with no proposed edit.

    line 181
  18. Distinctive:

    If the file no longer exists, explain its replacement or removal and do not propose recreating it without evidence.

    line 183
  19. Distinctive:

    Do not pad findings.

    line 205
  20. Distinctive:

    Ask which actionable findings to apply: all, selected finding numbers, or none.

    line 244
  21. Distinctive:

    After the user approves:

    line 247
  22. Distinctive:

    Never edit a tag.

    line 250
  23. Distinctive:

    Do not add a changelog entry for documentation-only edits.

    line 256

“Common” is measured: kept by at least 6 in 10 of the 5 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

Docs Drift

Verify that documentation tells the truth about either the changes about to ship or the latest version already shipped.

Keep the audit read-only. Do not edit, commit, tag, push, publish, or release until the user explicitly approves specific findings. Even after approval, never commit or publish from this skill.

Invocation

  • $docs-drift or $docs-drift unreleased — audit every plugin with a non-empty [Unreleased] section.
  • $docs-drift <plugin-slug> — audit that plugin's unreleased changes.
  • $docs-drift latest — audit the latest reachable release of every tagged plugin.
  • $docs-drift latest <plugin-slug> — audit only that plugin's latest release.

Use unreleased mode by default. Treat "latest version", "latest release", or equivalent wording as latest-release mode.

Typical release order, when the sibling skills are installed:

$release-status → $pre-release-review → $docs-drift → $release

Step 1 — Scope plugins and snapshots

Never mix released and unreleased windows in one claim set.

Unreleased mode

Identify plugins whose [Unreleased] section contains non-blank content:

for p in plugins/*/CHANGELOG.md; do
  slug=$(basename "$(dirname "$p")")
  awk '/^## \[Unreleased\]/{f=1; next} /^## \[/{f=0} f && NF' "$p" | rg -q '.' \
    && echo "$slug"
done

If the user named a plugin, limit scope to that slug after verifying it exists and has unreleased content. If nothing is in scope, stop with:

No unreleased changes in any plugin — nothing to drift-check.

Use the working tree as the documentation snapshot. Keep incidentally noticed older drift separate as pre-existing debt.

Latest-release mode

Start from all plugins/*/CHANGELOG.md slugs, or only the user-named slug. Resolve each plugin's refs independently:

EMPTY_TREE=$(git hash-object -t tree /dev/null)

for p in plugins/*/CHANGELOG.md; do
  slug=$(basename "$(dirname "$p")")
  latest=$(git describe --tags --abbrev=0 --match "${slug}--v*" HEAD 2>/dev/null || true)

  if [ -z "$latest" ]; then
    printf '%s\n' "$slug: skipped (no reachable release tag)"
    continue
  fi

  previous=$(git describe --tags --abbrev=0 --match "${slug}--v*" "${latest}^" 2>/dev/null || true)
  printf '%s\n' "$slug latest=$latest previous=${previous:-<empty-tree>}"
done

Use only reachable <slug>--v* tags. Do not select an unreachable tag merely because its version sorts later. If a plugin has only one reachable tag, use the empty tree as its previous snapshot. If every scoped plugin lacks a tag, stop with:

No reachable plugin release tags — nothing to latest-release drift-check.

For each plugin, use:

  • claim window: <previous-tag>..<latest-tag>, or empty tree to latest tag;
  • changelog claims: the exact ## [<latest-version>] section at the latest tag;
  • documentation snapshot: the repository exactly at that plugin's latest tag.

Later commits and working-tree changes are not release claims in this mode. Different plugins may have different latest-tag snapshots; preserve that mapping throughout the audit.

Step 2 — Extract doc-checkable claims

In unreleased mode, read only each scoped plugin's working-tree [Unreleased] section.

In latest-release mode, derive the version from <slug>--v<version> and read the matching changelog section from the tagged snapshot:

version=${latest##*--v}
git show "$latest:plugins/$slug/CHANGELOG.md" | awk -v ver="$version" '
  index($0, "## [" ver "]") == 1 {f=1; next}
  /^## \[/ && f {exit}
  f
'

Also inspect the plugin-scoped release diff to catch doc-checkable consumer-visible changes that a changelog bullet may group:

release_diff_base=${previous:-$EMPTY_TREE}
git diff "$release_diff_base" "$latest" -- "plugins/$slug/"

Use that diff only in latest-release mode and only for claim discovery; this skill does not replace changelog-versus-code review.

Turn each claim into zero or more checks:

  • Removal or rename — a skill, command, script, hook, config key, doc file, or CLI changed identity.
  • New operator-visible surface — a new skill, command, config block, hook, CLI, or channel behavior should appear where siblings are enumerated.
  • Changed default or behavior — documentation may still state the previous default or behavior. Treat this as the highest-severity class.
  • Count or version claim — a hand-written number or version marker may still contain the old value.
  • Promised documentation — a cited file or anchor must exist and cover what the changelog promises.

Mark pure internal fixes, refactors, and test-only changes as no doc surface. Do not manufacture a finding for every bullet.

Step 3 — Sweep the matching documentation

For each scoped plugin, consider the files that exist in its selected snapshot:

  • plugins/<slug>/README.md
  • plugins/<slug>/docs/*.md
  • plugins/<slug>/CLAUDE.md
  • plugins/<slug>/AGENTS.md
  • top-level plugin extras such as SAFETY.md, DOCKER.md, or CONTRIBUTING.md

Also check relevant root documentation:

  • root README.md
  • root CLAUDE.md
  • root AGENTS.md
  • root documentation extras relevant to a claim

Work claim-first, not file-first:

  • Unreleased: use rg -n against the working tree. Sweep root files once against the union of all claims.
  • Latest release: use git grep -n <pattern> "$latest" -- <paths> and git show "$latest:<path>". Sweep root files at each plugin's own latest-tag snapshot. Deduplicate identical root findings in the report without losing the plugin/tag evidence.

Never silently substitute current documentation for a released snapshot.

When more than two plugins are in scope and agent delegation is available, delegate one plugin audit per worker. Give each worker that plugin's claims, mode, and snapshot ref, and require:

snapshot:file:line — contradicted claim — current text — suggested fix

Keep the root-file sweep in the main session because it needs cross-plugin deduplication. Respect the available agent-slot limit and use additional waves if needed. If delegation is unavailable, perform the checks sequentially.

In latest-release mode, re-check each snapshot finding against the current working tree:

  • If it is still stale, propose an edit to the current file.
  • If it is already corrected, report it under Fixed after release with no proposed edit.
  • If the file no longer exists, explain its replacement or removal and do not propose recreating it without evidence.

Step 4 — Judge meaningfulness

Propose an edit only when a reader would be factually misled:

  • Meaningful: references a removed or renamed surface; states an old default or behavior; gives a wrong count or version; omits a new surface from a sibling enumeration; or promises documentation that does not exist or does not cover the claim.
  • Not meaningful: tone, formatting, absent marketing coverage, missing documentation for internal fixes, or changelog accuracy that $pre-release-review owns.
  • Borderline: reader impact is unclear. Report it separately with a one-line rationale and no proposed edit.

This marketplace ships as a research preview. Removed operator surfaces are deleted from documentation. Do not propose a Legacy section, leftover config keys in examples, "legacy configuration" labels, or dual current/old descriptions. State the present surface only.

Do not pad findings.

Step 5 — Report and wait for approval

Use this shape:

# Docs Drift — <date>

## Scope
Mode: <unreleased | latest release>
In scope: <slugs with snapshot refs>
Skipped: <slugs and reasons>
Claim windows: <[Unreleased] | previous-tag..latest-tag per plugin>
Documentation snapshots: <working tree | latest tag per plugin>

## Findings
### <n>. <slug> — <misleading | stale-reference | missing-doc>
- Claim: <abbreviated claim and changelog/diff evidence>
- Drift: <snapshot>:<doc file:line> — currently says: "<excerpt>"
- Current tree: <still stale | changed | file removed>
- Proposed edit: <exact replacement text, or "add section X covering Y">

## Fixed after release (no edit proposed)
- <snapshot>:<file:line> — <what was stale and where it is now corrected>

## Borderline (no edit proposed)
- <snapshot>:<file:line> — <one-line rationale>

## No doc surface
- <briefly grouped internal or test-only claims>

## Pre-existing drift (not this window)
- <anything noticed outside the selected claim window>

## Verdict
<N> actionable findings across <M> current files | Docs are clean for this window.

Ask which actionable findings to apply: all, selected finding numbers, or none. The audit request itself is not approval to edit.

After the user approves:

  1. Apply only the approved edits to the current working tree with apply_patch. Never edit a tag.
  2. Match each document's existing style. Keep edits succinct: prefer one sentence or one table-cell correction, and do not restate a full policy when its established term or canonical link is enough. Delete a removed surface; do not relabel it as leftover, compatibility, or legacy support.
  3. Re-read the changed passages and run git diff --check.
  4. Do not add a changelog entry for documentation-only edits; repository guidance explicitly excludes them.
  5. Report the uncommitted files and verification. Never commit, push, tag, publish, or release from this skill.

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.