audit-loop

Category: Development & Coding | Uploader: minghinmatthewlam | Downloads: 0 | Version: v1.0（Latest）

Lightweight read-only audit. Both model families review evidence and findings. Iterate on disagreement or low confidence. No code changes.

Changelog: Source: GitHub https://github.com/minghinmatthewlam/agent-guards

Directory Structure

Current level: skills/audit-loop/

📁 references/
- 📄 review-protocol.md 47 B
📄 SKILL.md 1.7 KB

SKILL.md

---
name: audit-loop
description: "Lightweight read-only audit. Both model families review evidence and findings. Iterate on disagreement or low confidence. No code changes."
---

# Audit Loop

Read-only audit with cross-model verification. No code changes.

Planning before implementation → use `plan-loop` instead.
Reviewing completed code changes → use `review-loop` instead.

## Prerequisites

Read `~/dev/agent-guards/AGENTS.md` before starting — the safety and workflow conventions that govern all agent work.

## How It Works

1. **Investigate** the question. Collect evidence (files, commands, source material).
2. **Review** your findings with both model families per `references/review-protocol.md`. Spin up all review agents in parallel. Pre-warm any deferred tools (e.g. `ToolSearch` in Claude Code) before the first review round so cross-model calls don't fail.
3. **Iterate** if reviewers disagree, confidence is low, or key claims lack evidence.
4. **Write a final verdict**: scope reviewed, evidence basis, findings, residual risks, next steps.

Every meaningful claim must cite specific files, commands, or source material.

## Round Sizing

- R1: 1 native + 1 cross-model.
- R2+ only when triggered: confidence below 85, reviewer disagreement, or high-impact finding lacks evidence. Focus on previous round's findings and unresolved gaps — not a fresh re-review.
- Escalate to 3+1 for high-risk or persistent disagreement.
- Cross-model cap: always 1 per round. Cross-model tools serialize, so a 2nd would wait for the 1st — add native reviewers instead for more coverage.

## Confidence

After each round, assess confidence 0-100 with top unknowns and what evidence is still missing:
- Below 85: consider another round.
- 85+: eligible to close.

Comments 0

Latest | Hot

Please login before commenting.

No comments yet. Be the first one!

audit-loop

Directory Structure

SKILL.md

Report

Notice