report-claim-audit
Skill3 starsChanged 53 days agoMITView on GitHub
Audit a status report, dispatch digest, or closeout handoff by re-deriving every claim from the artifact that owns it, accounting for what the report omitted, and pricing each decision so the operator can reply with a number instead of re-reading the repo. Use when a report is about to drive a decision and its numbers have not been independently recomputed. Not for writing the report, repairing what the audit finds, or reviewing a diff — an audit that edits its subject destroys its own evidence.
In this repository59 other files it gives its agents.
- AGENTS.mdAGENTS.md
- agent-harness-settings-paritySkill
- agent-operating-contract-evolutionSkill
- ai-coding-agent-delegationSkill
- arch-improvement-planSkill
- contract-bound-deliverablesSkill
- cross-agent-repair-routingSkill
- define-goalSkill
- delegated-work-acceptanceSkill
- design-an-interfaceSkill
- design-mdSkill
- discord-rich-messageSkill
- dotnet-repo-workflowSkill
- evidence-backed-analytical-writingSkill
- evidence-boundary-disciplineSkill
- find-nearbySkill
- gddSkill
- generated-runtime-repairsSkill
- ghc-searchSkill
- git-auto-sync-pipelinesSkill
- github-code-reviewSkill
- github-repo-managementSkill
- grounded-citationsSkill
- harness-engineeringSkill
- hermes-context-file-loadingSkill
- hermes-modelsSkill
- hermes-updatesSkill
- hoa-introspect-distillSkill
- hoa-introspectSkill
- hoa-skill-reviewSkill
- hoa-skill-workSkill
- html-artifactSkill
- jupyter-live-kernelSkill
- kanban-workflowsSkill
- m365-copilot-cdpSkill
- measurement-campaign-evidenceSkill
- microsoft-officeSkill
- notes-to-prdSkill
- ocr-and-documentsSkill
- official-document-verificationSkill
- omp-headless-orchestrationSkill
- ops-claim-verificationSkill
- package-release-verificationSkill
- plan-to-issuesSkill
- project-phase-closeoutSkill
- research-stream-consolidationSkill
- review-revision-candidatesSkill
- rss-feedsSkill
- skill-library-curationSkill
- skill-porting-and-publishingSkill
- spec-convergenceSkill
- structural-validity-gatesSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- teams-cdpSkill
- tmux-agent-controlSkill
- whisperSkill
- writing-plansSkill
- writingSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 6 are distinctive, and 0 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.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 38
Verify absence claims with the negative search.
- Distinctive: line 41
When the report names a validation command, locate the script before trusting the result.
- Distinctive: line 64
Check git status and the worktree yourself.
- Distinctive: line 78
Keep recommendations inside the frozen envelope.
- Distinctive: line 83
Do not invent the surface you found missing.
- Distinctive: line 89
Never commit or push from an audit run.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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.
Files
Report claim audit
A daily dispatch, teammate handoff, or closeout digest is a set of claims, not evidence. This skill is the pass that turns it into a decision brief the operator can answer with numbers instead of re-reading the repo themselves.
0. Pin your authority before reading anything
Open the role contract that governs the run: who may write which paths, what needs operator
approval, what must never be written. Everything below is shaped by it — an audit run usually
licenses reading and at most a notes/STATUS write, never a repair of what it found.
1. Recount every claim from the artifact that owns it
- Numbers come from the per-item JSON, the control manifest, the index, or the validator's own output — never from the report's table. A report correct eight times running is still a hypothesis on the ninth.
- Re-derive derived counts (N of M complete, X cells missing a key) by walking the per-item files in code, not by re-reading the report's prose. Print what you computed next to what it claimed.
- Verify absence claims with the negative search. A confirmed absence is itself a finding about the surface: "no STATUS.md exists anywhere" reframes the run from reading status to deriving it, and that recurring cost belongs in the brief.
- When the report names a validation command, locate the script before trusting the result. A path that does not exist is your error, not a failed check: search the repo for the basename, run it, report that rc.
- Re-run the repo's cheap gates once (validator, test runner, portability check) and quote real counts — "205 tests, 0 failures, 8 environment skips" is proof; "tests pass" is not.
2. Look for what the report did not say
- Account for the deliverable set by diffing lists, never by grepping dump text. Enumerate the
paths inside the control manifest and the acceptance record, enumerate disk by artifact pattern,
diff both directions. A substring search over a serialized JSON dump misses a name nested under
another key and false-positives on hex hashes. Shape:
d = json.load(open('control/artifact-manifest.json')) recorded = set(re.findall(r'[A-Za-z0-9._/-]+\.md', json.dumps(d, ensure_ascii=False))) # diff against a glob of the deliverable filename schema - A file on disk but invisible to the record is a stronger finding than the "missing review" usually reported: it sits in a surface declared closed, so nobody will review it and later readers trust it. Lead with it. Read the accepted-artifact list separately from the manifest's general artifact list (rejected vs accepted decides the remedy) and count mentions in the closeout report — zero mentions corroborates the diff, it does not contradict it. Report four numbers together: on disk N, in manifest M, in accepted set K, mentioned V.
- Check
git statusand the worktree yourself; "clean tree" is a claim like any other, and it also tells you whether another writer is mid-edit. - Separate already ruled from pending. An approved exception (a known dangling-link baseline, a registered legacy class) is not a decision item — listing it again trains the operator to ignore the whole list. Name it in one line as covered by the standing exception.
3. Price the decisions
- Compute the cheap mechanical facts (per-item sizes, key counts, aggregate characters) even when you cannot rank quality: they convert "72 drafts" into "72 drafts = 3.7M characters of semantic review", which usually answers the question by itself.
- One take per decision: recommend first, then the mechanism that makes the rejected branch expensive (what must be re-opened, re-sealed, re-reviewed, or re-run). State plainly what you cannot infer at all (external filing numbers, operator-only facts) instead of hedging around it.
- Keep recommendations inside the frozen envelope. Proposing a new research round, a new evidence class, or a re-opened lifecycle is an operator decision, not an audit finding.
4. Act only inside your boundary
- Do not invent the surface you found missing. A new status/notes file in a repo that has none, or in a subset of its projects, reproduces exactly the unaccounted-file defect you are reporting; inside a manifest-bound directory it is silently wrong.
- Route surface gaps to the role that owns that path. Hand over the judgement, not a prescription: the facts, the two possible rulings, the boundary (no commit, no claim change, no new file surface). A teammate told "fill the gap" fabricates entries to satisfy you.
- Never commit or push from an audit run. Leaving the tree untouched is a reportable result — say it was deliberate, and why (writes need approval).
5. Output shape
- Verdict on the report: which claims you reproduced, with the numbers you computed.
- Findings the report missed, defect first.
- Numbered decisions, each with your take and the cost of the branch you reject, so the operator replies with a number instead of prose.
- What you did and did not touch, including delegated handoffs by name.
Plain text in a terminal, one point per line, no markdown decoration: the operator may be reading this on a phone.
Pitfalls
- Reporting the digest's own counts back as verification of the digest.
- Declaring a named validator clean from an assumed path, or unclean because the path was wrong.
- Substring-grepping a serialized manifest to prove a file is or is not listed; walk the structure and extract path-shaped strings instead.
- Keeping state "durable" by dropping a notes file into a sealed artifact directory.
- Answering a decision list with questions instead of a recommendation plus the reason to reject the costly branch; "it depends" hands the question back unchanged.
- Re-raising items the corpus already covered with an approved exception.
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
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.

