agentleFS
Iabstergo1Iabstergo1/pdf-to-study-kb/.claude/skills/kb-qa/SKILL.md

kb-qa

Skill2 starsChanged 10 days agoMITView on GitHub

Run QA / audit / coverage checks on the published knowledge base or pre-save candidates, producing a report and Review-Queue proposals. Use when the user says "run a KB QA / audit coverage / spot-check evidence / run the Q-chain / check for concept pollution". Semantic-health words (L4, contradiction, Q2 added value) belong to wiki-lint-semantic and must not be hijacked here.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Run broad QA over the published vault or a pre-save candidate: coverage, evidence spot-checks, formula screenshot spot-checks, concept …

    line 8
  2. Very specific:

    Read: wiki/index.generated.md, wiki/concepts/_registry.yaml, and the relevant source/concept/topic/comparison/synthesis/lesson pages.

    line 20
  3. Very specific:

    If the finding targets an adopted legacy page, the handoff must say that revise-adopted can authorize the later edit but cannot validate …

    line 58
  4. Very specific:

    Use outcome: "findings" with findings: [{id: "F1", reason: "specific issue and evidence"}].

    line 104
  5. Very specific:

    After a confirmed repair, start a new batch with parent_id equal to the original batch ID and pages listing affected pages.

    line 110
  6. Distinctive:

    Do not promise complete factual correctness.

    line 116
  7. Distinctive:

    If the model exits without successful structured registration, the Web marks this source review incomplete.

    line 119
  8. Distinctive:

    A sampling PASS is not a clean bill for the page: When you spot-check assertions rather than reading the page whole, the verdict covers …

    line 161
  9. Distinctive:

    Every Q-chain item has a question, evidence, judgement, and a follow-up action.

    line 178
  10. Distinctive:

    No wiki content page was edited directly.

    line 182

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “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

kb-qa — post-publish / pre-save QA report

Run broad QA over the published vault or a pre-save candidate: coverage, evidence spot-checks, formula screenshot spot-checks, concept pollution, an ljg-qa-style Q-chain. Produces a report and Review-Queue proposals only; it does not edit content pages.

1. Triggers / Non-triggers

  • Triggers: "run a KB QA", "audit coverage", "spot-check evidence", "run the Q-chain", "check for concept pollution", "pre-save QA".
  • Non-triggers: "semantic health check", "find contradictions", "does the comparison cover the key dimensions", "Q2 added value" belong to wiki-lint-semantic; handling an existing Review-Queue uses kb-review; read-only Q&A uses kb-query; a new source uses ingest. kb-qa and wiki-lint-semantic are mutually exclusive on triggers.

2. Inputs

  • Scope: whole vault, a domain, a source, a query-session, or a proposed write candidate.
  • Read: wiki/index.generated.md, wiki/concepts/_registry.yaml, and the relevant source/concept/topic/comparison/synthesis/lesson pages.
  • May read: pipeline-workspace/query-sessions/<run_id>/, wiki/Review-Queue/, the deterministic lint result.

3. Outputs

  • Published-source reviews use content-review below: one immutable SQLite review batch, reproducible Markdown + JSON under pipeline-workspace/reports/kb-qa/<run_id>.*, and individual review_proposals rows committed with the results. These are model results, not human acceptance. Legacy YAML history is read alongside them, never backfilled.
  • pipeline-workspace/reports/kb-qa/<run_id>.md: the QA report.
  • Unregistered/pre-save scopes may still write wiki/Review-Queue/kb-qa-<YYYY-MM-DD>.md proposals; label them as report-only, without claiming registered page coverage.
  • The report covers scope, the Q-chain, sampled items, findings, risk level, and the suggested follow-up skill (usually kb-review).
  • For an actionable finding on an adopted/published legacy page, distinguish the later revise-adopted write authorization from content attribution: record official evidence for the finding, but do not add a book to source_refs or claim that the legacy page has acquired a source.

4. Dependencies

  • CLI: python scripts/pipeline.py status; if needed python scripts/pipeline.py lint --source <source_id>. revise-adopted belongs to the confirmed-fix path in kb-review, never to this read-only QA run.
  • Protocols: docs/skill-runtime/schema.md, save-back-policy.md, concept-resolution.md.
  • Triggers mutually exclusive with wiki-lint-semantic: semantic-health words are not handled here.

5. Persisted artifacts

  • pipeline-workspace/reports/kb-qa/<run_id>.md.
  • wiki/Review-Queue/kb-qa-<YYYY-MM-DD>.md (only when an actionable finding exists).
  • The spot-check list: sampled pages, evidence refs, and Q-chain conclusions recorded in the report.

6. CLI commands

python scripts/pipeline.py status
python scripts/pipeline.py lint --source <source_id>

This skill never promotes, rolls back, or edits content pages; fixes go to kb-review or the matching write skill. If the finding targets an adopted legacy page, the handoff must say that revise-adopted can authorize the later edit but cannot validate the evidence or change source_refs.

7. Workflow

Published-source registration

Review one explicit source per batch. Whole-vault/domain requests can produce separate source batches; do not infer report ownership from today's shared-page source_refs. The Web provides a review tool with {action, value}; other agents use the same CLI: python scripts/pipeline.py content-review --action <action> --payload '<JSON>' (or --stdin). No direct SQLite writes or edits to content-review-ledger.yaml.

  1. begin: {run_id: UUID, source, method: "full-page" | "sampled-claims", pages?: [vault-relative paths]}. Full-page defaults to every currently published content page of this source. For sampling, explicitly select pages. The returned scope is frozen; its denominator counts knowledge pages, not original PDF pages. The Web supplies source/run/method.
  2. read: {run_id, path} for every selected page at the time you review it. The tool returns exact content and a stored read_id, revision and time. Read the original text/images too. Ordinary file reads alone do not supply review receipts. A full-page check examines every substantive assertion; sampling must name only the assertions actually checked. Read receipts are evidence of tool delivery, not proof of comprehension.
  3. register: submit the following object after every scoped page has been checked. In Web pass it as value to review. The CLI validates it, creates both report formats from this structured result, and atomically registers results and findings. Do not additionally create duplicate Markdown proposals for these findings.
{
  "format": 1,
  "run_id": "<UUID from begin>",
  "summary": "Scope, method, limits and overall observations",
  "pages": [{
    "path": "domains/example/concepts/example.md",
    "read_id": "<receipt from read>",
    "outcome": "no-findings",
    "checks": [{
      "claim": "Exact assertion checked, or the explicit full-page assertion inventory",
      "evidence": [{"path": "pipeline-workspace/staging/<source>/source.md", "locator": "page/section/paragraph"}],
      "judgment": "Evidence-based conclusion and limits"
    }],
    "findings": []
  }]
}

Use outcome: "findings" with findings: [{id: "F1", reason: "specific issue and evidence"}]; IDs must be unique within the batch. Use inconclusive when evidence cannot establish a conclusion. Evidence references can also name source images or official HTTPS citations. Missing evidence or incomplete work must not become no-findings. Retrying the identical registration is safe; changing a registered result is refused.

  1. After a confirmed repair, start a new batch with parent_id equal to the original batch ID and pages listing affected pages; review their new contents with new receipts. The Web recheck action supplies these fields. Old findings stay in history; neither recheck nor lint closes them automatically. Their disposition is a separate human action.
  2. status shows batch results and changed versions. Never call accept or close on behalf of the user. They inspect reports/dispositions and confirm the batch in Web. Do not promise complete factual correctness. Reading all knowledge pages does not prove every statement in the original source was captured.

If the model exits without successful structured registration, the Web marks this source review incomplete. Continue the same job to finish it. Do not re-hash repaired pages at registration time: the receipt version is always the reviewed version.

Sub-unitInputOutputAcceptancePersistedFailure stop
QA1 scoperequest + indexscope and samplesdoes not hijack wiki-lint-semantic triggersreport draftscope unclear
QA2 status backgroundsource/status/lintdeterministic backgrounddoes not re-implement deterministic lintreportlint already blocked
QA3 Q-chainin-scope materialquestion→evidence→judgement→actionevery Q has an evidence pathreportevidence thin
QA4 spot-checkpage/formula/evidence samplesspot-check resultssamples and conclusions traceablereportsample missing
QA5 file findingsactionable findingsReview-Queue proposalno content-page editskb-qa proposalduplicate proposal

Evidence discipline (two rules that cost real rework when broken)

A "not in the source" verdict requires an exhaustive search, and the search must be shown. Before writing that a term/number/mechanism is absent, search the source for the Chinese rendering, the English name, the abbreviation, and case/hyphen variants — then record every search term in the finding. A single-language search is the documented way this goes wrong: an audit once declared a bilingual term "not mentioned" after searching only its English form, while the book used the Chinese form several times. The cost is asymmetric — a missed finding leaves one clause unfixed, but a false finding triggers a rewrite of a correct page, and every rewrite is a fresh chance to introduce an error.

Normalize five PDF extraction deformations before you conclude a term is absent. A page's word can be verbatim in the source yet fail a naive string search because the extracted source.md mangled it. One audit hit the first three at once (20 false "not in source" candidates), then the full page-by-page read surfaced two more — each of which, if believed, would have rewritten a correct page. Apply all five before the search, not after:

  1. Ligatures — the PDF stores ffi/ff/fi/fl (single code points, e.g. U+FB03) so Efficiency is Efficiency, Parameter-Efficient is Parameter-Efficient. Expand ligatures to their ASCII letters.
  2. Line-break hyphenation — a word split across a line break keeps its hyphen: Accurately becomes Ac-⏎curately, Prefixes becomes Pre-⏎fixes. Rejoin -⏎ before searching.
  3. CJK–Latin spacing — the extractor inserts a space between CJK and Latin runs, so Encoder-only架构 is stored as Encoder-only 架构. Collapse all whitespace (this also covers the plain Chinese/English/abbreviation/case/hyphen variants above).
  4. Thousands separator — the page writes 1000 but the source prints the grouped form 1,000 (ROME's "1000 条知识" is 1,000 in the source). Strip , from digit runs before matching numbers.
  5. Chinese numerals — the page writes an Arabic digit but the source uses the Chinese numeral (T5's "连续 3 个" is 连续三个 in the source). Try the 一/二/三… form of any small integer.

Record the normalized form you searched, not just the raw term.

A sampling PASS is not a clean bill for the page. When you spot-check assertions rather than reading the page whole, the verdict covers only the assertions actually sampled — say so in the report, and later rounds must never inherit it as evidence that the page is clean. Three consecutive rounds on one source overturned pages that an earlier sampling round had marked clean; only an integral, per-assertion pass supports a page-level conclusion.

8. Failure stops / recovery

vault/index missing; scope unclear; the request is really a semantic-health check (switch to wiki-lint-semantic); deterministic lint already failed with no stable QA scope; an evidence path is missing; the user asks to edit content pages directly (switch to kb-review and wait for confirmation); an adopted-page finding has no official evidence strong enough to support the proposed correction. Recovery: the report + proposals are the durable output; re-run after the scope is clarified.

9. Acceptance criteria

  • The QA report is written under pipeline-workspace/reports/kb-qa/.
  • Every Q-chain item has a question, evidence, judgement, and a follow-up action.
  • Actionable findings are written as Review-Queue proposals.
  • Adopted legacy findings separate “evidence supports this correction” from “the pipeline authorizes a later write”; no new source_refs are inferred from a citation or from revise-adopted.
  • No wiki content page was edited directly.
  • None of wiki-lint-semantic's exclusive triggers were handled.

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.