kb-lint — Claude Code skill: what it does, how to install
Audits an agent knowledge bundle for specification conformance and drift. Safely repairs minor link and metadata defects upon request.
npx skills add stjbrown/agent-knowledge --skill kb-lint -a claude-codeWhat kb-lint does
The skill inspects target bundles such as knowledge/ in two distinct passes. First, it runs a zero-dependency script at scripts/conformance.mjs against frontmatter metadata. The check requires Node.js 18 or newer. The script checks compliance with SPEC §11, flags broken links, and validates ISO date formats.
Next, the skill performs a drift audit. It looks for orphaned concepts with no incoming links, outdated statements, and unresolved contradictions across notes. It also verifies types against spec/types.md and detects broken references to repository files. It categorizes findings into Error, Warning, and Info tiers. A dated summary goes to log.md.
Running the tool with fix applies safe repairs. It regenerates stale overview notes and cleans up obvious link targets. The skill leaves claim meanings untouched. Resolving factual contradictions or modifying schema types always requires manual confirmation.
What it helps with
- verify bundle frontmatter against SPEC §11 using Node.js
- detect orphaned concepts that lack inbound links
- surface contradictions and outdated statements across notes
- auto-fix broken links and malformed log dates safely
- append structured audit records to log.md
Example prompts for the agent
- Check the
knowledge/bundle for conformance and drift - Audit all concept files and show any SPEC §11 errors
- Run
kb-lint fixto correct broken links and stale overviews
How to install kb-lint
1Skills CLI (Vercel)
npx skills add stjbrown/agent-knowledge --skill kb-lint -a claude-code2Manually: put the files in place
~/.claude/skills/kb-lint/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/kb-lint/
1Skills CLI (Vercel)
npx skills add stjbrown/agent-knowledge --skill kb-lint -a codex2Manually: put the files in place
.agents/skills/kb-lint/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add stjbrown/agent-knowledge --skill kb-lint -a cursor2Manually: put the files in place
.cursor/skills/kb-lint/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install stjbrown/agent-knowledge kb-lint2Manually: put the files in place
.github/skills/kb-lint/SKILL.md1Skills CLI (Vercel)
npx skills add stjbrown/agent-knowledge --skill kb-lint -a gemini-cli2Platform CLI
gemini skills install https://github.com/stjbrown/agent-knowledge --path skills/kb-lintManually: put the files in place
.devin/skills/kb-lint/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add stjbrown/agent-knowledge --skill kb-lint -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about kb-lint
How do I install kb-lint in Claude Code?
Run npx skills add stjbrown/agent-knowledge --skill kb-lint -a claude-code to register the skill in your project.
What does kb-lint do?
It evaluates a structured knowledge bundle for syntax compliance and semantic drift. It detects stale claims, orphaned concepts, broken links, and schema divergences.
What issues can kb-lint auto-fix?
In fix mode, it rebuilds stale overviews from child notes, repairs obvious broken links, and standardizes log dates. It never alters the meaning of factual claims without explicit approval.
Source text
SKILL.md · MIT
---
name: kb-lint
description: Health-check a knowledge bundle for conformance and drift; optionally auto-fix safe issues.
disable-model-invocation: true
version: 0.3.2
tags: [knowledge, okf, lint, conformance]
---
# kb-lint — health-check the bundle
Keep a [bundle](../kb/SKILL.md) trustworthy as it **compounds** by catching **drift** — the decay a
growing knowledge base accumulates. Two passes: a **deterministic conformance** check (mechanical,
scripted) and a **drift audit** (fuzzy, judgment). Run both; report findings by severity. With
`fix`, repair what is safe.
## 1. Conformance (deterministic)
For a bundle-wide format or lifecycle migration, first inventory active metadata with the same
frontmatter-aware checker:
```
node "<skill-dir>/scripts/conformance.mjs" <bundle-dir> --inventory --json
```
The inventory parses only each concept's leading YAML frontmatter and reports body `# Citations`
separately. Its `observations` include counts and exact paths for every active top-level key and
status value; `legacy_citations` counts citation sections, numbered items, separately defined
footnote sources, total source records, and common citation shapes; `structured_sources` counts the
postflight concepts, entries, and entries carrying `resource`. An absent observation means zero. Do
not use repository-wide grep to classify active metadata because prose and fenced examples can contain
inactive field names and lifecycle values. Resolve every reported frontmatter error before planning
from the counts.
Run the bundled checker against the target bundle (default `knowledge/`). It is a zero-dependency
Node script (`node >=18`); `<skill-dir>` is this skill's directory — `${CLAUDE_SKILL_DIR}` under
Claude Code, or whatever path your host exposes for the skill:
```
node "<skill-dir>/scripts/conformance.mjs" <bundle-dir>
```
It reports **ERROR** for a hard [SPEC](../kb/references/SPEC.md) §11 failure and **warn** for
producer-profile defects in optional v0.2 fields, broken links, non-ISO log dates, and stale dates.
Optional-family warnings never make the bundle non-conformant. Broken links are explicitly
tolerated by §6.1.
**Completion criterion:** the checker has run and every ERROR it reported is listed for the report
(and fixed, if in `fix` mode).
## 2. Drift audit (judgment)
The checker can't see meaning. Audit the bundle for the ways a compounding artifact rots — this is
the legwork that makes lint worth running. Cover every check:
- **Contradictions** — concepts asserting conflicting facts that aren't linked `conflicts_with`.
- **Stale claims** — statements a newer source has superseded but that were never marked
`superseded_by`; overviews behind their children.
- **Orphans** — concepts with zero inbound [cross-links](../kb/references/glossary.md) (index/log
exempt; overviews exempt).
- **Missing cross-references** — concepts about the same entity/theme that don't link to each other.
- **Coverage gaps** — entities named repeatedly across concepts but lacking their own concept; data
gaps a source or web search could fill.
- **Provenance gaps** — concepts making external claims with neither OKF v0.2 `sources` nor a legacy
`# Citations` / Reference trail.
- **v0.2 metadata drift** — `sources` entries without `resource`, footnote labels without a matching
`sources[].id`, malformed `generated`/`verified` actors or dates, invalid lifecycle values, and
Attested Computations missing their required runtime or computation contract.
- **Repository drift** — for living repository documentation, missing paths in
`sources[].resource`, evidence that no longer supports the documented behavior, or source changes
since `spec/repository_state.md`'s `documented_revision` that affect a concept without a
corresponding update. A lingering `documented_worktree: true` there requires comparison with the
current working tree. Skip this check when repository evidence or Git history is unavailable.
- **Schema drift** — types used but absent from `spec/types.md`; documented types that no longer
describe their concepts; spelling/case variants; or one overloaded type hiding several recurring,
materially distinct entity kinds. Treat unused documented types as Info, not an error.
**Completion criterion:** every check above has been run across the whole bundle and its findings
recorded — not a sample.
## 3. Report
Present findings grouped by check, each tagged:
- **Error** — §11 conformance failures. The bundle is non-conformant until fixed.
- **Warning** — drift that degrades trust (contradictions, stale claims, orphans, broken links).
- **Info** — suggestions (coverage gaps, new concepts or sources worth adding).
Turn coverage gaps into concrete next moves: questions to investigate, sources to
[ingest](../kb-ingest/SKILL.md), or repository evidence to inspect through
[kb-document](../kb-document/SKILL.md). Append a dated summary (counts + notable findings) to the
bundle's `log.md` — append-only.
**Completion criterion:** a severity-grouped report is delivered and a `log.md` summary appended.
## Fix mode
If invoked with `fix`, repair only what is **safe and unambiguous**, then re-report what was fixed
vs. what needs a human:
- **Safe to auto-fix:** stale overviews (regenerate from children), missing cross-links, malformed
log dates, broken links with an obvious target, index entries out of sync with files.
- **Never auto-fix:** anything that changes a claim's meaning. A contradiction or a stale *claim* is
resolved by [ingest](../kb-ingest/SKILL.md) under the [trust model](../kb/references/trust-model.md)
(**supersede**/**conflict**) — never by editing meaning in place here. Type renames, merges,
splits, deprecations, and migrations also require user confirmation; report the proposed schema
change and affected concepts together. Route living repository behavior that needs a meaning
update to [kb-document](../kb-document/SKILL.md), which applies the trust model's narrow
versioned-repository exception with source evidence and a revision log.
**Completion criterion:** every safe issue is fixed and every meaning-level issue is flagged (not
touched); the re-report distinguishes the two.
© stjbrown · MIT · stjbrown/agent-knowledge