Agentic Audit

Multi-agent adversarial verification pipeline. Three independent agents examine content from orthogonal dimensions, then the host agent rewrites based on findings — delivering the corrected output directly, not the audit report.

Pipeline

Trigger word ("audit"/"review"/"审计"/"审查")
  → Launch 3 agents in parallel
    ├── verifier-depth  (completeness)
    ├── verifier-bias   (logic & reasoning)
    └── verifier-fact   (factual accuracy)
  → Host receives all 3 reports
  → Rewrite content incorporating findings
  → If ≥ 5,000 characters → generate DOCX + open
  → If < 5,000 characters → output markdown only

Key Design Rules

1. Deliver the product, not the process

Don't show audit reports and ask "should I proceed?" — the user said "audit" to get a corrected result, not to see intermediate findings. Rewrite directly. If the user wanted step-by-step, they would have asked for it.

2. External monitoring over self-monitoring

A single agent cannot reliably detect its own errors (self-correction blind spot: ~64.5% failure rate). This skill uses three independent external agents for verification. The host agent acts on their findings, not on its own self-assessment.

3. Decision heuristics over value derivation

In execution paths (when triggered), use fast pattern-matching decision heuristics. Keep the "why" in annotation position — it's for understanding off-trigger, not for reasoning during execution. This prevents motivated reasoning from converting "why" into violation rationalizations.

4. Rule conflict resolution

When two instructions/tendencies conflict:

1. User's explicit one-step instruction > conversational default posture
2. Product delivery obligation > process display comfort
3. Affirmative behavioral encoding > diffuse personality tendency

Never present the conflict to the user ("memory says A but I think B, which do you choose?") — resolve it yourself.

Agent Specifications

verifier-depth

Checks for missing dimensions: genealogies, broken logic chains, unexamined root problems, alternative perspectives, key distinctions unmade.

Agent definition: agents/verifier-depth.md

verifier-bias

Checks for reasoning errors: confirmation bias, coherence traps, pseudo-certainty, sycophancy/flattery, logical jumps.

Agent definition: agents/verifier-bias.md

verifier-fact

Checks factual claims: distinguishes verified facts, inferences, uncertainties, and errors. Cites sources.

Agent definition: agents/verifier-fact.md

Output Format

The rewritten output should:

• Start by correcting any factual errors found
• Then present the revised analysis
• End with a summary table of what changed from the original
• Clearly label remaining uncertainties and irreducible failure modes

Boundaries

• The 5,000-character threshold for DOCX generation is estimated by Chinese character count (including punctuation, excluding code blocks)
• If audit reveals significant factual errors affecting the core argument, briefly inform the user before the rewrite
• Audit reports are tools, not products — they don't need the user's attention unless they contain information the user MUST know before seeing the rewrite

References

• Design principles and known limitations: references/design-principles.md
• For implementation in Claude Code: use UserPromptSubmit hook (not Stop hook — broken for prompt type as of v2.1.81). Alternatively, use Dynamic Workflow for native parallel agent support.