English

I built this because running everything on the most expensive model is a double waste: it burns tokens on grunt work a cheaper model does fine, and — worse — the model that wrote the code is way too nice grading its own homework. The idea comes straight from Addy Osmani's Loop Engineering:

"The most useful structural thing in a loop, by far, is splitting the one who writes from the one who checks."

So this skill makes your top model (Fable, or whatever you run the main loop on) act as the commander: it drafts the plan, asks you which model should do the research, reviews what comes back, asks you which model should execute, and then runs a maker/verifier loop where the checker is always a fresh, independent context.

One model doing everything vs. the commander workflow

The workflow

Phase 0  Commander drafts Plan v0 (acceptance criteria + known-unknowns)
Phase 1  Ask the user: which model for research?
Phase 2  Research fan-out on that model (facts + sources only, no plan edits)
Phase 3  Commander reviews → Plan v1 (trust but verify)
Phase 4  Ask the user: which model for execution?
Phase 5  Execution loop: maker (chosen model) ⇄ verifier (fresh context)
Phase 6  Wrap-up: final audit + adoption-rate record + lessons fed back

Built-in guardrails: iteration cap of 3 rounds per gate, escalate to human after 2 rounds without improvement, a hard human gate before anything irreversible (commit / push / publish / spend), and an explicit ban on the verifier rejecting work for style or opinion — it can only flag those as non-blocking suggestions.

Install

mkdir -p ~/.claude/skills/fable-commander
cp SKILL.md ~/.claude/skills/fable-commander/SKILL.md

Then say: "用指揮官模式跑這個題目" / "run this in commander mode".

Note: the skill text is written in Traditional Chinese (the SKILL.md is the product — a fully specified workflow). Claude follows it regardless of the language you converse in.

See it in action

• examples/PLAN-sample.md — a real-shaped state file after a full Phase 0–6 run: acceptance criteria, gate-by-gate rounds, adoption rate.
• examples/verifier-report-sample.md — what a verifier report looks like: PASS/FAIL with evidence, plus the "suggestions (non-blocking)" lane that protects your style.

When not to use it

The skill checks this itself (Phase −1): if the task is small, the verification can't be made objective, or the token budget is tight — it will tell you to just do the task directly. Loop engineering is not a universal answer.

Related

Sister project: loop-engineering-reviewer — run your loop with the commander, then audit the loop itself against the seven Loop Engineering rules.

Credit

The structural beliefs (maker/verifier split, objective gates, stop points, adoption rate as the success metric) come from Addy Osmani's Loop Engineering. This skill is one opinionated way to put them to work inside Claude Code.

繁體中文

做這個 skill 的起點是一個挫折：把所有事都丟給最貴的模型，是雙重浪費 —— 查資料、寫樣板這種粗活燒它的 token 很不划算；更糟的是，寫程式的模型自己驗收自己，永遠改作業改得太仁慈。理念直接來自 Addy Osmani 的《Loop Engineering》：

「迴圈裡最有用的結構性設計，就是把『寫的人』和『查的人』分開。」

所以這個 skill 讓你的最強模型（Fable，或任何跑 main loop 的模型）只當指揮官：起草計畫 → 問你研究要派哪個模型 → 審查研究產出 → 問你執行要派哪個模型 → 跑 maker/verifier 分離的執行迴圈，驗收永遠由獨立 fresh-context 的 agent 來做。

一個模型做到底 vs 指揮官工作流

流程總覽

Phase 0  指揮官起草 Plan v0（驗收標準 + 已知未知清單）
Phase 1  問使用者：研究階段用哪個模型？
Phase 2  研究 fan-out（該模型）：只回報事實與來源，不改 plan
Phase 3  指揮官 review → Plan v1（trust but verify）
Phase 4  問使用者：執行階段用哪個模型？
Phase 5  執行迴圈：maker（執行模型）⇄ verifier（獨立 fresh context）
Phase 6  收尾：總驗收 + 採納率記錄 + 教訓回灌

內建護欄：每個 gate 迭代上限 3 輪、連續 2 輪無改善就升級人工、不可逆動作（commit / push / 發布 / 花錢）前一律過人工閘、明文禁止 verifier 以風格或觀點理由打回——那些只能寫進「建議（不擋交付）」欄。

安裝

mkdir -p ~/.claude/skills/fable-commander
cp SKILL.md ~/.claude/skills/fable-commander/SKILL.md

之後說：**「用指揮官模式跑這個題目」**即可觸發。

實際長什麼樣

• examples/PLAN-sample.md —— 一份跑完 Phase 0–6 的範例 state file：驗收標準、逐 gate 輪次紀錄、採納率。
• examples/verifier-report-sample.md —— verifier 報告長相：PASS/FAIL + 證據，以及保護你風格的「建議（不擋交付）」欄。

什麼時候不該用

skill 自己會先檢查（Phase −1）：任務太小、驗證無法客觀化、或 token 預算緊——它會直接告訴你「這題直接做比較划算」。Loop engineering 不是萬用解。

姊妹作

loop-engineering-reviewer —— 用 commander 跑迴圈，再用 reviewer 拿七條 Loop Engineering 量尺回頭體檢迴圈本身。

致謝

結構信念（maker/verifier 分離、客觀 gate、停止點、採納率作為成功指標）全部來自 Addy Osmani 的《Loop Engineering》。本 skill 只是把它落地成 Claude Code 裡一套有立場的工作流。

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。

📧 [email protected] · 💼 LinkedIn