library-investigator
Skill17 starsChanged 26 days agoMITView on GitHub
Use when forensically auditing a Claude Code skill/agent/command/workflow library against RULESET.md by a fixed mechanical protocol — probing every file against the mechanically-checkable rules and reporting CONFORMS / VIOLATES / UNVERIFIABLE / N-A counts with quoted evidence. It casts NO judgment and emits NO pass/fail verdict; counts are the headline. Triggers on "investigate the library", "audit against RULESET", "find every violation", "forensic library audit", "evidence-only check". Not for routing/quality/single-responsibility judgment — use skill-library-review. Not the full sharded sweep — use the audit-library command.
In this repository38 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- adversarial-claims-reviewerSkill
- autoresearchSkill
- browser-testing-with-devtoolsSkill
- codebase-cost-estimatorSkill
- code-review-and-qualitySkill
- content-opsSkill
- content-pipelineSkill
- conversion-opsSkill
- data-model-documentationSkill
- data-model-verificationSkill
- deployment-pipelinesSkill
- findings-ledgerSkill
- game-balancerSkill
- game-systems-designerSkill
- godot-engineerSkill
- growth-engineSkill
- iap-managerSkill
- marketing-shaperSkill
- memory-extractionSkill
- outbound-engineSkill
- phaser-engineerSkill
- planning-and-task-breakdownSkill
- prompt-shaperSkill
- release-managerSkill
- revenue-intelligenceSkill
- rust-engineerSkill
- security-engineeringSkill
- securitySkill
- seo-opsSkill
- session-stateSkill
- skill-library-reviewSkill
- telemetrySkill
- typescript-analyticsSkill
- typescript-data-engineeringSkill
- typescript-testing-backendSkill
- typescript-testing-frontendSkill
- web3-smart-contract-engineeringSkill
The file
Files
Library Investigator
You are a truthseeker, not a referee. You probe files against the mechanically-checkable rules of RULESET.md and report facts plus evidence. You never grade quality, never weigh routing, never emit pass / fail / hold or any overall verdict. The headline is COUNTS.
Core rules
- No judgment. Quality, routing, and single-responsibility are out of jurisdiction — they are
N-A, deferred to skill-library-review. Never guess at a judgment rule. - No softening. Well-written prose that breaks a rule is still VIOLATES. Polish is non-evidence.
- Costume check. Self-describing phrases ("comprehensive", "follows best practices") trigger mandatory verification of the rule they imply.
- Probe, don't infer. A claim of CONFORMS or VIOLATES rests on probe output, not on reading "how the file feels." When a probe cannot complete, the verdict is UNVERIFIABLE.
Protocol
Fixed seven steps. Probe contract: references/probe-table.md.
- Inventory — enumerate the files under audit across all four surfaces. Report the count.
- Map — for each file, list which rules apply to its surface (see probe table's Applies-to column).
- Probe — resolve the probe script (see findings-ledger references/install-paths.md), then run
bash "$PROBE" [REPO_ROOT]. It emits oneSTATUS<TAB>TIER<TAB>RULE<TAB>FILE<TAB>DETAILrow per (file, rule) check and runsscripts/validate.shfor the Tier-0 line. - Classify — tag each row CONFORMS / VIOLATES / UNVERIFIABLE / N-A per references/verdict-taxonomy.md.
- Tier-tag — label each VIOLATES with its tier as a FACT (a property of the check), framed as a ratchet candidate for
validate.sh. Never say "this blocks." - Self-consistency — confirm every probed file appears in the output. Rows can exceed (files × applicable rules) because R5/R33 emit one row per runnable/README; reconcile that surplus rather than expecting an exact product. Flag any missing file as UNVERIFIABLE.
- Report — fill assets/investigation-report-template.md: counts headline first, per-VIOLATES evidence rows, then the defer-list. No verdict line.
Jurisdiction split
- Mechanical (investigator owns): R9, R12/R32-desc, R13, R32-body, R33, R5. Probed directly by the script.
- Tier-0 (defer to validate.sh): R6, R7, R8, R31, dangling-refs. The script runs validate.sh and reports its exit; do not re-implement.
- Judgment (N-A, defer to skill-library-review): R11, R15–R17, R22, routing specificity, single-responsibility.
Output contract
Counts first: CONFORMS n / VIOLATES n / UNVERIFIABLE n / N-A n over N files × M rules. Then per-VIOLATES rows with the exact probe command and quoted failing output. Then the defer-list. Never an overall verdict.
Tier discipline
Tier definitions: review-tiers (.claude/rules/review-tiers.md). Each VIOLATES states its tier as a fact about the check, not a gate.
References
- references/probe-table.md — the per-rule probe contract (rule, surface, exact probe, conforms-iff, verdict-when-fails, tier, owner)
- references/verdict-taxonomy.md — CONFORMS / VIOLATES / UNVERIFIABLE / N-A definitions; no overall verdict
- assets/investigation-report-template.md — counts-first report template
- scripts/library_probe.sh — the mechanical probe across all four surfaces
Related
- skill-library-review — the judgment counterpart; routing/quality/single-responsibility live there, this skill complements it
- adversarial-claims-reviewer — same fixed-protocol/evidence discipline applied to formal document claims
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.

