agent constitution audit

AI agents fail when their governing rules contradict each other or drift from the core constitution. This skill performs a rigorous, read-only audit of stacked context layers, including identity files, global instructions, memory stores, and skill directories to find these fractures.

What it does

• Inventory layers by locating identity files and instructions using real precedence orders.
• Connection check to verify if your written standards are actually being referenced or ignored.
• Conflict scanning to find pairwise contradictions across layers with specific quoted evidence.
• Memory hygiene to identify bloated stores, leaked secrets, or rules that weaken agent safety.
• Behavioral dry-runs to test agent judgment against safety prompts and injected instructions.

How it works

1. Point the auditor at your agent's home directory or configuration files.
2. The agent inventories all active layers and checks for broken references or duplicate rules.
3. It performs a read-only scan of your persistent memory and skill triggers.
4. You receive a structured report with PASS/FAIL statuses and specific evidence for every claim.

Frameworks & tools

Works with any agent using rule-layer systems like CLAUDE.md, .cursorrules, SOUL.md, or custom MCP skill directories.

Why this beats prompting it yourself

Manual reviews often miss cross-file contradictions or "hidden" instructions in memory stores. This skill uses a systematic pairwise comparison and evidence-based reporting that forces the agent to prove every status with a file path or direct quote rather than hallucinating consistency.

Use cases

• Auditing a new agent setup to ensure it follows company-wide coding standards.
• Debugging why an agent has stopped following specific instructions.
• Checking persistent memory for privacy leaks or "safety-weakening" entries.
• Validating that newly added skills don't have overlapping triggers.

Known limitations

This is a read-only audit and will not apply fixes automatically. Some behavioral tests may return UNKNOWN if the agent's internal reasoning cannot be fully exposed.