audit-team
Skill0 starsChanged 4 months agoMITView on GitHub
Forensic audit of any AI agent team. Checks agents, skills, protocols, memory, comms, and tool integration for gaps, staleness, contradictions, and missing wiring. Use when auditing a team, reviewing agents, checking for gaps, forensic review, or verifying a build is ready.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- build-agentSkill
- build-pipelineSkill
- build-skillSkill
- build-teamSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 2 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
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
- Distinctive: line 13
Use these as the quality standard:
- Very specific: line 27
Multi-division team: if divisions/*/CLAUDE.md exists -> expect memory at divisions/*/memory/.
- Very specific: line 28
Single-team: if no divisions/ directory -> expect memory at memory/ at root.
- Very specific: line 48
If a team mixes patterns (e.g., says "Engineering division" in CLAUDE.md but has no divisions/ directory), flag as STALE terminology.
- Distinctive: line 99
Run these specific checks:
- Very specific: line 101
Skill-CLAUDE alignment: Verify every skill listed in CLAUDE.md has a corresponding file in .claude/skills/.
- Very specific: line 102
Agent-division alignment: Verify every agent listed in division CLAUDE.md has a file in .claude/agents/.
- Very specific: line 104
Authority consistency: Verify VETO/HALT/BLOCK holders in CLAUDE.md match those declared in agent definitions.
≈ “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.
Files
Audit Team
Trigger
"Audit the team", "review the agents", "check for gaps", "forensic review", "are the agents ready?", "is everything wired correctly?", "verify the build".
Research (Reference During Audit)
Use these as the quality standard:
standards/floor-spec.md-- canonical floor anatomy. Score the team against its §Conformance Checklist (20 items: CLAUDE.md size/stamp/map, roster parity, authority tiers, file ownership, skills location, activation, comms freshness, memory caps, handoff hygiene, lifecycle gates) in addition to the per-file checks below.standards/skill-authoring.md-- skill rules (frontmatter, 200-line cap, no absolute paths, changelog)research/agent-architecture-playbook.md-- agent definition standards, comms requirements, quality controlresearch/skill-design-mastery.md-- skill sizing, archetype compliance, design rule violationsresearch/team-design-blueprint.md-- org pattern validation, handoff system, red team coverage
Process
Step 1: Inventory
List every file in the system using Glob. First determine the team's structure pattern, then verify presence of the files that pattern requires.
Structure detection:
- Multi-division team if
divisions/*/CLAUDE.mdexists -> expect memory atdivisions/*/memory/. - Single-team if no
divisions/directory -> expect memory atmemory/at root.
Common requirements (both patterns):
CLAUDE.md(workspace level)- All agent
.mdfiles in.claude/agents/ - All skill
.mdfiles in.claude/skills/ - Protocol files in
protocols/(minimum: escalation-matrix, session-close) - Comms files:
comms/wire.md,comms/board.md,comms/channels/all-hands.md - Rules files in
.claude/rules/(minimum: desk-protocol) -- optional for small single-team setups .mcp.jsonif team uses external tools
Multi-division specifics (only if divisions/ exists):
- All division
CLAUDE.mdfiles indivisions/*/ - Memory files in
divisions/*/memory/(minimum: MEMORY.md, learning-ledger.md) - Per-division channel in
comms/channels/[division].md
Single-team specifics (only if no divisions/ directory):
- Memory file at
memory/MEMORY.md - Single domain channel in
comms/channels/[team-domain].md(e.g.,dev.md,sales.md)
If a team mixes patterns (e.g., says "Engineering division" in CLAUDE.md but has no divisions/ directory), flag as STALE terminology.
Step 2: Per-File Checks
Agent definitions (grep for key sections):
- Has YAML frontmatter with name, description, model, tools
- Mandate is specific and non-overlapping (grep for file ownership conflicts)
- Red team agents have
disallowedTools: [Write, Edit, Bash] - First Action reads division memory
- Autonomous Failure Triggers section exists
- Dependency map has upstream AND downstream
- Output format is structured
- Under 100 lines
Skills (validate against 8 design rules):
- Description has 3+ trigger phrases
- Pre-reads list specific file paths
- Process has numbered steps with binary gates (not advisory)
- Output format specifies save location
- Under 200 lines (300 for pipelines)
- References correct file/function names (not stale)
- No hardcoded absolute paths (skill-authoring rule)
- Circuit breaker defined
- Authority tier declared (ADVISORY/COMMIT/BLOCK/VETO/HALT)
Protocols:
- Escalation matrix covers L1-L4 with specific triggers
- Session close has a checklist
- No contradictions between protocol files
Memory:
- MEMORY.md under 150 lines
- learning-ledger.md exists with format template
- No references to killed/archived items as active
Communication:
- Wire has format header and priority levels defined
- Board has format header and status values defined
- Every division has a channel
- kill-log.md exists (or will exist on first kill)
- skill-changelog.md exists (or will exist on first self-edit)
MCP/Tool Integration (if applicable):
.mcp.jsonexists with valid server configurations- Each server has dedicated venv (command points to venv python)
- No hardcoded secrets in
.mcp.json(use env vars) - Tool count per server under 20
- Agent definitions document which MCP tools they use and when
Step 3: Cross-File Consistency
Run these specific checks:
- File ownership conflicts: Grep all agent definitions for "Files Owned" and verify no path appears in two agents.
- Skill-CLAUDE alignment: Verify every skill listed in CLAUDE.md has a corresponding file in
.claude/skills/. - Agent-division alignment: Verify every agent listed in division CLAUDE.md has a file in
.claude/agents/. - Broken references: Grep all files for file paths and verify referenced files exist.
- Authority consistency: Verify VETO/HALT/BLOCK holders in CLAUDE.md match those declared in agent definitions.
Step 4: Report
For each file, assign a status:
| Status | Meaning |
| PASS | No issues |
| STALE | Content outdated (cite specific item) |
| CONTRADICTS | Conflicts with another file (cite both) |
| MISSING | Expected content absent (what should be there) |
| ORPHAN | File exists but nothing references it |
| OVERSIZED | Exceeds line cap |
Summary table:
| File | Status | Issue | Fix |
|------|--------|-------|-----|Step 5: Grade
| Grade | Criteria |
| A | 0 issues. Ship it. |
| B | 1-3 issues, all STALE or cosmetic. Ship with fixes queued. |
| C | 4-10 issues or 1+ CONTRADICTS or MISSING. Fix before operating. |
| D | 10+ issues or critical wiring gaps (no comms, no protocols). Major rework. |
| F | System non-functional. Missing CLAUDE.md, no agents, or no skills. Rebuild. |
Issue severity:
- CONTRADICTS or MISSING authority/ownership = counts as 3 issues
- STALE date or cosmetic = counts as 1 issue
- ORPHAN = counts as 0.5 issues (not harmful, just messy)
Parallel Fan-Out Note
For any multi-item batch (N ≥ 2 posts, audits, reviews, analyses, cleanups), spawn one top-level agent per item, running in parallel, rather than routing the batch through a single orchestrator that processes items serially or via internal sub-agents invisible to the user.
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.

