Korean | 한국어

, 반복되는 중립 QA 대리인 실행을 위한 휴대용 신속한 본 루프. 그것은 파일을 검토 할 수있는 코딩 에이전트와 함께 작동, diffs, 로그, 테스트 산출, 또는 생성한 artifacts는, 복잡한 vibe-coded 프로젝트를 지킵니다 비스듬한 자기 검사에 붙어있는 대신 새로운 버그를 서핑.

문제

AI 코딩 에이전트는 자체 검토에 유용하지만 나중에 디버깅 패스 종종 덜 효과적입니다. 에이전트가 이전 가정을 본 경우, 의심스러운 수정, 또는 이전 결론, 그것은 같은 검사를 계속하는 경향이 지역 및 버그의 다른 클래스를 놓습니다.

복잡한 프로젝트를 위해, 더 강한 본은 반복한 QA를 가진 통과하는 것을 입니다 대화 역사가없는 신선한 에이전트. 각 통행은 단지 받습니다 본래 필요조건, 검열하는 구체적인 artifacts, 및 권위 옵션 정보 선명한 의심의 여지 없이 검토자를 선도해야 합니다. 예상 결과.

이 제공하는 것

• 독립적 인 QA 에이전트를위한 중립 검토 신속한 구조.
• 원래 목표, 리뷰 대상 및 컨텍스트 소스 사이의 단단한 분리.
• 의심스러운 버그를 방지하는 규칙, 수정, 이전 의견, 또는 예상 결론.

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