agentleFS
noahcostakesnoahcostakes/claudebrain-template/.claude/skills/audit-extreme/SKILL.md

audit-extreme

Skill0 starsChanged 5 months agoMITView on GitHub

EXHAUSTIVE sibling of /audit. Deep-reviews EVERY file (notes, raw data, config, AND binaries/PDFs), web-verifies EVERY externally-checkable claim, and surfaces every finding with a per-file manifest. Read-only — never edits your notes. Expensive: always shows a plan and confirms before running. Invoke as /audit-extreme <folder> [focus text].

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 5 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use it when you want maximum certainty and accept the token/time cost.

    line 12
  2. Distinctive:

    If the path is a single file → mode = SINGLE-FILE.

    line 24
  3. Very specific:

    If the path is a protected/excluded area (_meta/, 40-archive/, attachments/, .obsidian/, .git/) → warn and require explicit confirmation.

    line 25
  4. Very specific:

    Read shared context ONCE (same as /audit): the folder's README.md/CLAUDE.md, _meta/orient.md, _meta/tags.md, _meta/profile.md if user-facts …

    line 30
  5. Distinctive:

    ALWAYS print a REVIEW PLAN and wait for explicit confirmation before dispatching:

    line 32
  6. Very specific:

    If zero in-scope files → report *"Nothing to audit in <folder>."* and stop.

    line 34
  7. Distinctive:

    Write the report file. This is the ONLY file the skill writes.

    line 67
  8. Distinctive:

    READ-ONLY by default. Agents get no edit/write permission.

    line 73
  9. Distinctive:

    ALWAYS print the Phase 0 review plan and get confirmation before dispatching — extreme is expensive.

    line 74
  10. Very specific:

    Never auto-add tags. Tag-vocabulary violations are reported as suggestions for the user to approve in _meta/tags.md.

    line 75
  11. Very specific:

    Use [[wikilinks]] in the report and chat summary, never markdown links.

    line 82

≈ “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

Audit-Extreme Skill

What this does

The maximum-thoroughness version of /audit. Reads and deep-reviews EVERY in-scope file regardless of type — notes, raw/generated data (validated and cross-checked against the notes), config/scaffold, and binaries (PDFs/images read via the Read tool) — and web-verifies EVERY externally-checkable claim (not just uncertain ones). Surfaces every finding, with a full per-file manifest. Writes ONE tiered report — _audit-extreme-YYYY-MM-DD-HHMM.md — inside the reviewed folder. READ-ONLY: never modifies reviewed files; offers confirmed fixes afterward.

Use it when you want maximum certainty and accept the token/time cost. For routine, fast reviews use /audit instead.

Arguments

  • /audit-extreme <folder-path> [focus instructions]
  • Same as /audit: <folder-path> is relative to vault root; optional trailing FOCUS DIRECTIVE is threaded to every agent; no path → default to the current note's parent folder and confirm.

Steps

Phase 0 — Map & plan (ALWAYS confirm before dispatching)

  1. Resolve the folder path relative to vault root. If it does not exist, report and suggest closest-matching folders (Glob); stop.
  2. If the path is a single file → mode = SINGLE-FILE; review just that file (still full depth + full fact-check).
  3. If the path is a protected/excluded area (_meta/, 40-archive/, attachments/, .obsidian/, .git/) → warn and require explicit confirmation.
  4. Enumerate every file (Glob <folder>/**/*). Classify each FOR THE MANIFEST, but review ALL classes deeply:
    • KNOWLEDGE NOTE, RAW/GENERATED DATA, CONFIG/SCAFFOLD, BINARY/ATTACHMENT — ALL get deep review.
    • HARD-SKIP only: .git/ internals, lockfiles (*.lock, package-lock.json, etc.), and prior _audit-*.md / _audit-extreme-*.md reports.
    • Any file that cannot be read (corrupt/opaque binary) → record "could not read" in the manifest; never silently drop.
  5. Read shared context ONCE (same as /audit): the folder's README.md/CLAUDE.md, _meta/orient.md, _meta/tags.md, _meta/profile.md if user-facts are in scope, and notes wikilinked from key files.
  6. Choose mode: ≤ 3 in-scope files → SINGLE-PASS (coordinator reviews inline); otherwise FAN-OUT with SMALL chunks (~5–10 files/agent, because each file gets deep review + per-claim web checks); cap concurrency ~6–10 agents.
  7. ALWAYS print a REVIEW PLAN and wait for explicit confirmation before dispatching:
    • file counts by class, total in-scope files, number of agents, and the warning: "Extreme audit reads every file (including binaries) and web-verifies every claim — this is token- and time-intensive."
    • If zero in-scope files → report "Nothing to audit in <folder>." and stop.
    • On user decline → stop cleanly; suggest narrowing scope or using /audit.

Phase 1 — Review (read-only, EXHAUSTIVE)

  1. FAN-OUT: dispatch one read-only agent per chunk IN PARALLEL (a single message with multiple Agent calls); use a read-only agent type (Explore or general-purpose) and tell it explicitly NOT to edit or write any file. SINGLE-PASS / SINGLE-FILE: the coordinator reviews inline.
  2. Each agent's BRIEF — review EVERY assigned file at full depth across ALL lenses:
    • FACTUAL ACCURACY — flag wrong claims AND mark EVERY externally-checkable claim verify: web (not just uncertain ones) for Phase 2 batch verification.
    • INTERNAL CONSISTENCY — math, dates, within-file contradictions, contradictions vs. the shared context.
    • DATA INTEGRITY (raw/generated data) — validate syntax (JSON/XML/CSV parses), schema/shape sanity, suspicious/missing/duplicate values; and CROSS-CHECK data values against the notes that cite them (e.g. a course's credit hours in the data vs. in a note).
    • CONFIG CORRECTNESS (config/scaffold) — review CLAUDE.md / .claude/ skills & agents / settings for correctness, broken references, and stale instructions.
    • BINARY CONTENT (PDFs/images) — read via the Read tool; summarize and flag anything wrong or inconsistent with the notes.
    • VAULT HYGIENE — frontmatter valid (type, date, status, tags); tags ∈ _meta/tags.md; wikilinks for internal refs (flag internal markdown links; external URLs via markdown are fine); source: where derived.
    • WRITING QUALITY — clarity, structure, redundancy, dead placeholders.
    • LOCAL STRUCTURE — naming/placement oddities.
    • Return STRUCTURED FINDINGS ONLY: {file, location, severity: Critical|Important|Minor, confidence: High|Med|Low, lens, issue, suggestion, verify: web?} PLUS a one-line per-file verdict for the manifest (clean / issues found / could-not-read). Do not modify any file.
  3. Collect all findings, all verify: web claims, and the per-file verdicts.

Phase 2 — Synthesis & FULL verification

  1. Whole-folder checks (same as /audit): broken wikilinks; cross-note contradictions; duplicates; structure (orphans, misfiled, missing-but-referenced, naming); tag-vocab rollup (flag non-approved tags as suggestions, never auto-add).
  2. WEB VERIFICATION — verify EVERY verify: web claim with WebSearch/WebFetch. Record each as confirmed / wrong / unverifiable WITH a source URL — including confirmed ones. Batch related claims. If web tools are unavailable, mark every such claim "unverified — no web access"; never guess.
  3. De-duplicate findings; assign final severity + confidence; group by file; assemble the per-file manifest.

Phase 3 — Report (surface EVERYTHING)

  1. Compute <folder>/_audit-extreme-<YYYY-MM-DD-HHMM>.md. Read templates/audit-report.md and substitute:
    • {{date}} → now (YYYY-MM-DD HH:MM); {{scope}} → folder path; {{mode}} → extreme (N agents) / extreme single-pass / extreme single-file; {{focus}} → focus directive or —.
    • {{verdict}} → 1–3 sentence overall health + headline issues.
    • {{scorecard}} → one row per dimension exercised (Factual / Consistency / Data integrity / Config / Binaries / Hygiene / Structure / Writing): | <dimension> | ✅|⚠️|❌ | <count> |.
    • {{critical}}, {{important}}, {{also-noted}} → bulleted findings (file + location, what's wrong, fix, evidence). Include ALL minor/low-confidence items — suppress nothing. Empty section → —.
    • {{fact-check-log}} → EVERY web-verified claim (confirmed AND wrong) → verdict + source URL.
    • {{coverage}} → FULL per-file manifest: every in-scope file with its verdict (clean / issues / could-not-read), plus a note of any hard-skipped files. Include file and agent counts.
  2. Write the report file. This is the ONLY file the skill writes.
  3. Post a chat summary: verdict + scorecard + the top issues + a [[_audit-extreme-…]] wikilink to the full report.
  4. Offer: "This was read-only — I changed nothing. Want me to fix any of these? I'll confirm each change before writing." If yes, fix ONE BATCH AT A TIME with confirmation, honoring the Rules below.

Rules

  • READ-ONLY by default. Agents get no edit/write permission. The coordinator writes exactly ONE file: the report. Reviewed files are never modified.
  • ALWAYS print the Phase 0 review plan and get confirmation before dispatching — extreme is expensive; never skip this gate.
  • Never auto-add tags. Tag-vocabulary violations are reported as suggestions for the user to approve in _meta/tags.md.
  • Cite a source URL for every externally-verified fact. Mark anything unverifiable as such — never fabricate a citation.
  • Skip prior _audit-*.md / _audit-extreme-*.md reports, .git/ internals, and lockfiles when enumerating.
  • Warn + confirm before auditing _meta/, 40-archive/, attachments/, .obsidian/, .git/.
  • The report note uses type: audit, tags: [audit]. Never tag it #review.
  • Degrade gracefully: if web/MCP tools are unavailable, complete the audit and mark affected items "unverified" rather than failing or guessing.
  • Any post-report fixing confirms each change, never deletes without explicit confirmation, and never edits _meta/profile.md or _meta/memory.md without permission.
  • Use [[wikilinks]] in the report and chat summary, never markdown links.

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.