skill
loop-engineering-reviewer
Loop Engineering 合規審查員 — 用 Addy Osmani《Loop Engineering》七條量尺審 AI 迴圈專案（agent/skill/cron），保留風格只修合規。Claude Code skill + subagent，中英雙語。
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/DennisWei9898/loop-engineering-reviewer itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026loop-engineering-reviewer
A compliance reviewer for AI loop projects — audit against Loop Engineering, then fix while keeping your style. AI 迴圈專案的合規審查員 —— 用 Loop Engineering 量尺體檢，再保留你的風格只修合規。
Review your agent / skill / cron loop against seven Loop Engineering rules, then fix only the structure — never your voice. 用七條 Loop Engineering 量尺審你的 agent / skill / cron 迴圈，只修結構、不碰你的語氣。
English
A reviewer built on Addy Osmani's Loop Engineering. Use it to produce and review an AI loop project (agent / skill / cron automation). It does two things:
- Review — audits your project against seven Loop Engineering rules, gives
PASS / FAIL / WARN+ evidence from the source, and points to what to change. - Fix — after confirming with you, edits directly but fixes only compliance and fully preserves your original style and opinions. A check-up and a prescription, not a transfusion.
Why you need it
As you wire more agents / skills / crons together, the thing that breaks isn't usually a feature bug — it's the loop structure:
- letting the writer grade its own homework (maker/verifier not separated)
- a gate that can't actually reject bad output ("take the highest of 3 rounds")
- a loop running unattended, failing unattended (Ralph Wiggum)
- an out-of-domain judge (a docs expert grading SEO claims)
This is the tool that hunts exactly those "loop structure diseases".
The 80/20 questioning discipline
It won't dump 50 questions on you. When it finds a pile of issues:
- it surfaces the 3–5 highest-impact questions first (real bug > gate weakness > safety > the rest)
- minor issues are fixed when encountered or batched later
- it raises compliance to 100% progressively, reporting
X/7 → Y/7each pass - if you insist on doing it all at once, it lists everything but confirms one at a time — never overloading you
Install (skill, subagent, or both)
As a Skill (Claude Code)
mkdir -p ~/.claude/skills/loop-engineering-reviewer
cp SKILL.md ~/.claude/skills/loop-engineering-reviewer/SKILL.md
Then say "use loop-engineering-reviewer to audit this agent."
As a Subagent (Claude Code)
cp loop-engineering-reviewer.agent.md ~/.claude/agents/loop-engineering-reviewer.md
Then spawn the loop-engineering-reviewer subagent to run an audit in isolation.
Skill = review + fix inline in the main chat. Agent = a fresh-context audit in its own session (closer to Loop Engineering's "verifier fresh context" ideal).
The rubric: seven Loop Engineering rules
| # | Rule | In one line |
|---|---|---|
| L1 | Maker/Verifier separation | writer ≠ checker; the verifier is an independent fresh context |
| L2 | Objective gate | pass conditions are machine-checkable / re-runnable, not "looks good" |
| L3 | Minimal viable loop, 4 parts | automation + skill + state file + a gate that can reject bad output |
| L4 | Stop point + iteration cap | round limits; no "take the highest score" softening |
| L5 | Human gate before irreversible | confirm before commit / push / publish / spend |
| L6 | Success metric = adoption rate | track how often output is accepted, not just tokens |
| L7 | Guard the three traps | silent failure / comprehension debt / cognitive surrender |
See it in action
examples/scorecard-sample.md — a full sample scorecard: L1–L7 verdicts with quoted evidence, the 80/20 question triage (3 questions first, not 50), and the "fix compliance, never your voice" proposal.
Related
Sister project: fable-commander — run your task with a commander workflow (plan → delegated research → maker/verifier loop), then audit the loop itself with this reviewer.
Credit
The rubric is entirely from Addy Osmani's Loop Engineering (also on O'Reilly Radar). This tool just lands it as a runnable review flow.
License
MIT — use, modify, and share freely.
繁體中文
一個建立在 Addy Osmani《Loop Engineering》上的合規審查員，用來幫你「產出並 review 一個 AI 迴圈專案」（agent / skill / cron 自動化）。它做兩件事：
- 審 —— 用七條 Loop Engineering 量尺逐條檢查你的專案，給
PASS / FAIL / WARN+ 原文證據，指出要改哪裡。 - 改 —— 跟你確認後直接動手修，但只修合規性、完整保留你的原始風格與觀點。體檢後開藥，不是換血。
為什麼你會需要它
你把 agent / skill / cron 越串越多之後，最容易壞的不是功能 bug，而是迴圈結構：
- 讓寫的人自己改自己的作業（maker/verifier 沒分離）
- gate 其實否決不了爛輸出（「3 輪取最高分」這種軟化）
- 迴圈無聲地一直跑、一直錯（Ralph Wiggum）
- 拿領域外的裁判去審（用文件專家審 SEO 主張）
這支就是專門抓這類「loop 結構病」的體檢工具。
八二法則提問紀律
它不會一次拋 50 個問題壓垮你。發現一堆問題時：
- 先挑影響最大的 3–5 個關鍵問題（真 bug > gate 命門 > 安全 > 其餘）
- 細碎的問題遇到再改或分批
- 循序漸進把合規度拉到 100%，每批回報
X/7 → Y/7進度 - 你若堅持一次到位，它會列全部，但逐一確認，不一次塞爆你的認知
安裝（skill、subagent，擇一或都裝）
當 Skill（Claude Code）
mkdir -p ~/.claude/skills/loop-engineering-reviewer
cp SKILL.md ~/.claude/skills/loop-engineering-reviewer/SKILL.md
之後說「用 loop-engineering-reviewer 審一下這個 agent」即可觸發。
當 Subagent（Claude Code）
cp loop-engineering-reviewer.agent.md ~/.claude/agents/loop-engineering-reviewer.md
之後 spawn loop-engineering-reviewer subagent 在獨立 context 跑審查。
Skill 版 = 在主對話裡順手審＋改。Agent 版 = 丟進獨立 context 跑一份審查（更貼近 Loop Engineering 的 verifier fresh context）。
量尺：Loop Engineering 七條
| # | 量尺 | 一句話 |
|---|---|---|
| L1 | Maker/Verifier 分離 | 寫的人 ≠ 審的人，verifier 獨立 fresh context |
| L2 | Gate 客觀可判 | 過閘條件是機器可判/可重跑，不是「看起來不錯」 |
| L3 | 最小可行迴圈四零件 | 自動化 + skill + 狀態檔 + 一道能否決爛輸出的關卡 |
| L4 | 有停止點 + 迭代上限 | 有輪次上限，禁止「取最高分放行」軟化閘門 |
| L5 | 不可逆前人工閘 | commit/push/發布/花錢前拿人類確認 |
| L6 | 成功指標＝採納率 | 記錄產出被採納的比例，不是只看 token |
| L7 | 防三大陷阱 | 無聲失敗 / 理解債 / 認知投降 各有對策 |
實際長什麼樣
examples/scorecard-sample.md —— 一份完整範例計分卡：L1–L7 逐條判定 + 原文證據、八二法則提問分流（先問 3 題，不是 50 題轟炸）、「只修合規不碰風格」的修正提案。
姊妹作
fable-commander —— 用指揮官工作流跑任務（規劃 → 委派研究 → maker/verifier 迴圈），再用本 reviewer 回頭體檢迴圈本身。
致謝
量尺全部來自 Addy Osmani 的《Loop Engineering》（O'Reilly Radar 亦有轉載）。本工具只是把它落地成一份可執行的審查流程。
授權
MIT —— 自由使用、修改、分享。
Contact · 合作聯絡
Open to collaboration on AI agent workflows. 對 AI agent 工作流的合作提案持開放態度，歡迎聯繫。
📧 dennis.xd.wei@gmail.com · 💼 LinkedIn