agent-self-audit

name	Agent Self Audit
description	Meta-auditing of Claude agents and their outputs. Use when: reviewing agent quality, validating agent decisions, or improving agent prompts.

Agent Self Audit

Meta-auditing framework for Claude agent quality.

Audit Dimensions

Dimension	Metrics	Target
Correctness	Bug rate	< 5%
Efficiency	Tokens/task	baseline ± 20%
Safety	Prompt injection defense	100%
Coherence	Follows instructions	> 90%

Self-Audit Checklist

1. Input Validation

• User intent correctly understood
• Edge cases handled
• Ambiguous requests clarified

2. Output Quality

• Matches user request
• Follows project conventions
• No hallucinated information

3. Code Quality

• TypeScript strict compliant
• No security vulnerabilities
• Tests written where needed

4. Safety

• No credential leakage
• Input sanitized
• Output validated

Evaluation Criteria

interface AgentEvaluation {
  task: string;
  intentMatch: 'exact' | 'partial' | 'missed';
  codeQuality: {
    types: boolean;
    security: boolean;
    tests: boolean;
  };
  efficiency: {
    tokens: number;
    timeMs: number;
  };
  issues: string[];
}

For Mansoni Agents

Audit questions:

1. Did the agent follow CLAUDE.md rules?
2. Did it use correct skills?
3. Did it maintain code quality standards?
4. Did it handle errors properly?

Continuous Improvement

Track metrics over time:

• Task success rate
• Token efficiency
• Common failure modes
• Skill usage effectiveness