fable-commander

Make your top model the commander, not the coder — it plans and reviews; the models you choose do the research and the work. 讓最強的模型當指揮官，而不是碼農 —— 它只規劃與審查，研究和執行派給你選的模型。

A Claude Code skill that turns your session into a commander workflow: plan → delegated research → review → maker/verifier execution loop. 一個 Claude Code skill，把你的 session 變成指揮官工作流：規劃 → 委派研究 → 審查 → maker/verifier 分離的執行迴圈。

English · 繁體中文

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.

指揮官派工手冊.pdf — a visual, walk-through handbook (Traditional Chinese): when a request is big enough to dispatch to the commander, how the two AskUserQuestion checkpoints work, and when the Advisor mode is actually the right call.

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.

Advisor mode: a lighter alternative

Claude Code also ships an experimental /advisor <model> command: your main model runs cheap, and Claude automatically consults a stronger advisor at key decision points (before committing to an approach, on repeated errors, before declaring done). Anthropic's own docs say the advisor fits long, multi-step tasks where plan quality determines the outcome, and adds less value on short tasks — for those, they recommend switching the main model instead. So the sweet spot inside Phase −1's escape valve isn't "the task is small" — it's when verification can't be made objective or the token budget is tight, but the task itself is still long-running or multi-step.

It's not a substitute for the commander workflow, though — the advisor gives a second opinion to the same executor, which usually just follows it; there's no independent verifier, no objective gate, no stop condition. It also has two known rough edges worth knowing before you rely on it: (1) if your session (or any subagent inside it) has loaded a deferred tool via ToolSearch — which stock Claude Code routinely does — advisor calls afterward deterministically fail (GitHub #73923, closed as not planned); (2) with Fable 5 as the main model, very long transcripts (~100K+ tokens) can silently disable the advisor too — this cap is specific to Fable 5 as the request model, not frontier models generally (the same report shows Opus 4.8 as the main model handling a 296K-token transcript fine) (GitHub #67609).

See the "Advisor 模式" section in SKILL.md for full guidance, including the trap of running this skill's verifier while an advisor is still attached — it erodes the independence the whole workflow depends on.

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。

指揮官派工手冊.pdf —— 圖文教學手冊：什麼需求夠格叫指揮官出馬、兩個 AskUserQuestion 卡點怎麼答、以及什麼時候該開 Advisor 模式。

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。

Advisor 模式：更輕量的替代方案

Claude Code 另外還有一個實驗性的 /advisor <model> 指令：主模型跑便宜的，Claude 在關鍵決策點（提交方案前、錯誤重複、宣告完成前）自動諮詢更強的 advisor。官方文件原文是 advisor 適合「長、多步驟」任務、方案品質決定成敗，短任務「加值較低」、官方建議短任務直接換主模型。所以 Phase −1 逃生口裡真正的甜蜜點，不是「任務小」，而是「驗證難以客觀化」或「token 預算緊」——但任務本身仍偏長偏多步驟。

但它取代不了指揮官工作流——advisor 是給同一個執行者的第二意見，執行者通常照做；沒有獨立 verifier、沒有客觀 gate、沒有停止條件。而且有兩個值得先知道的已知坑：(1) 如果你的 session（或任何 subagent）曾用 ToolSearch 載入過 deferred 工具——標準 Claude Code 常態性會這樣做——之後的 advisor 呼叫會確定性失敗（GitHub #73923，Anthropic 已標記不打算修）；(2) 用 Fable 5 當主模型時，很長的對話（約 100K+ tokens）也可能讓 advisor 悄悄失效——此限制專屬 Fable 5 當主模型，同一份回報顯示 Opus 4.8 當主模型處理 296K tokens 仍正常（GitHub #67609）。

完整說明見 SKILL.md 的「Advisor 模式」節，包含「開著 advisor 跑本 skill 的 verifier」這個陷阱——它會侵蝕整個工作流賴以運作的獨立性。

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