Autonomous QA Loop

English | 中文

A portable prompt pattern for running fresh, neutral QA agents in repeated loops. It works with coding agents that can review files, diffs, logs, test output, or generated artifacts, and helps complex vibe-coded projects keep surfacing new bugs instead of getting stuck in biased self-checks.

Problem

AI coding agents are useful at self-review, but their later debugging passes often become less effective. Once an agent has seen prior assumptions, suspected fixes, or earlier conclusions, it tends to keep checking the same areas and miss different classes of bugs.

For complex projects, the stronger pattern is to run repeated QA passes with fresh agents that have no conversation history. Each pass should receive only the original requirement, the concrete artifacts to inspect, and authoritative context. The prompt must avoid leading the reviewer toward known suspicions or expected findings.

What This Provides

• A neutral review prompt structure for independent QA agents.
• A hard separation between original goal, review target, and context sources.
• Rules that prevent leaking suspected bugs, intended fixes, prior opinions, or expected conclusions.
• A repeatable QA-loop pattern: split broad scopes by module, run fresh agents, triage the combined findings, then run another fresh neutral pass.
• Lightweight validation guidance that keeps QA focused on necessary core behavior instead of heavyweight audit gates or process-only safeguards.
• A plain prompt file plus an optional packaged version for local agent tooling.

Core Prompt

Use PROMPT.md with any agent. The generated QA prompt must contain exactly four top-level sections:

Background
Goal (Original Request)
Review Target
Relevant Context Documents

Run it as a loop: split broad scopes into module-level packets, send each packet to a fresh agent, triage both defects and plausible concerns in the main thread, fix confirmed issues, then start another fresh pass. Repeat until independent passes stop surfacing meaningful defects or concerns. Keep validation lightweight: cover necessary core behavior, and avoid audit gates or process-heavy safeguards unless the original goal or production safety requires them.

Optional Codex Skill Install

Copy the included skill folder into your Codex skills directory:

Copy-Item -Recurse .\autonomous-qa-loop $env:USERPROFILE\.codex\skills\

On macOS or Linux:

cp -R ./autonomous-qa-loop ~/.codex/skills/

Then start a new Codex session so the skill metadata is discovered.

Use it like this:

Use $autonomous-qa-loop to draft a neutral prompt for a fresh independent QA agent.

Repository Layout

.
├── PROMPT.md
├── README.md
└── autonomous-qa-loop/
    ├── SKILL.md
    └── agents/
        └── openai.yaml

License

CC0 1.0 Universal. You can copy, modify, redistribute, and use this prompt pattern for any purpose.

中文

这是一个通用的“自动 QA 循环”提示词模式，用来让全新的、无历史上下文的独立 agent 反复审查复杂代码变更。任何能看文件、diff、日志、测试输出或生成物的 coding agent 都可以使用。

它解决的问题

vibe coding 做复杂项目时，经常会堆出很多隐藏 bug。让同一个 AI 自查虽然能发现 一部分问题，但第二、第三轮 debug 往往会被前面的上下文、猜测和修复方向影响， 很难继续发现新的问题。

更有效的做法是每一轮都开启全新的、无历史上下文的 agent。大范围审查时先按模块 拆成多个中立审核包并行跑，只给每个 agent 原始需求、具体审核对象和权威上下文， 不告诉它已知怀疑、预期问题或前一轮结论。agent 需要报告明确缺陷和可疑风险点， 主线程统一判断哪些是真 bug。校验必须保持轻量，只覆盖必要的核心生产需求，避免 无生产价值的审计 gate、治理流程或重型校验管线。确认问题并修复后，再开启下一轮 新的中立审核。

内容

• PROMPT.md：通用提示词规则，可给任何 agent 使用。
• autonomous-qa-loop/：可选的本地 agent 工具包装。

Codex Skill 安装方式（可选）

Copy-Item -Recurse .\autonomous-qa-loop $env:USERPROFILE\.codex\skills\

macOS 或 Linux：

cp -R ./autonomous-qa-loop ~/.codex/skills/

使用：

Use $autonomous-qa-loop to draft a neutral prompt for a fresh independent QA agent.

许可证

CC0 1.0 Universal。你可以自由复制、修改、再发布，或用于任何目的。