Skill: system-prompt-audit

Purpose

Audit all system-level instructions embedded in this repository's Claude configuration,
surface them as a structured inventory, and validate them for consistency, safety,
and alignment with project goals.

Inspired by CL4R1T4S (Latin: "clarity") — the open-source project that collects and
publishes AI system prompts to promote transparency.

When to Use

• Run before merging changes to .claude/ directory
• Run as part of release readiness checks
• Run when a new agent or skill is added
• Run on-demand for governance audits

Steps

1. Inventory Collection

Scan the following paths:
  .claude/agents/       → agent persona definitions
  .claude/skills/       → skill behavioral specs
  .claude/commands/     → command implementations
  CLAUDE.md             → root-level project instructions

For each file, record:

• File path
• Word count (proxy for complexity)
• Presence of: role definition, constraints, examples, output format

2. Consistency Check

Cross-reference all files for:

• Duplicate roles: two agents claiming the same responsibility
• Contradictory constraints: one skill allows X, another forbids X
• Missing output formats: skills that don't specify what they produce
• Undocumented side effects: skills that write files/commit without stating so

3. Safety Check

Flag any instructions that:

• Allow unrestricted file system writes
• Allow network calls without validation
• Grant escalated permissions
• Lack human-in-the-loop checkpoints for destructive operations

4. Generate Audit Report

Write docs/system-prompt-audit.md with:

# System Prompt Audit Report
Generated: <timestamp>

## Summary
- Total components audited: N
- Agents: N | Skills: N | Commands: N
- Issues found: N (critical: N, warnings: N, info: N)

## Inventory Table
| Component | Type | Role | Has Constraints | Has Examples |
|-----------|------|------|-----------------|--------------|

## Issues
### Critical
...
### Warnings
...
### Info
...

## Full Behavioral Inventory
[per-component breakdown]

5. Exit Codes

• 0: No issues found
• 1: Warnings found (non-blocking)
• 2: Critical issues found (should block merge)

Integration

Add to .claude/commands/review.md as an optional step:

- [ ] Run system-prompt-audit before final review sign-off

Related Skills

• prompt-transparency — generates human-facing transparency docs
• release-readiness — broader release gate checks
• council-review — multi-agent review process