一套指揮一群 AI 寫程式代理人的兩階段工作流，作者本人不會寫程式。 A two-phase workflow for directing a swarm of AI coding agents — written by someone who does not write code.
中文 · English
核心想法：會不會寫程式不是重點。 重點是把「你要什麼」講到精確、可驗收， 再在 AI 動工時盯緊它別亂跑。這四個可組合的 skill，就是把這個想法變成能跑的東西。
緣起：「不會寫 code 的醫師，怎麼架構一批臨床專案」部落格文章。
| Skill | 層級 | 做什麼 |
|---|---|---|
/scope |
計畫 | 把一句話的需求拆成 N 張可驗收的契約（Agent Brief）。每條驗收標準都標上 [AFK]（agent 自己做自己驗）或 [HITL]（要你親自簽名）。完全不寫 code。 |
/execute |
執行 | 把一張契約從頭跑到底：一個 runner 寫、一個不同模型的 reviewer 審，最多退回 3 次，然後開 PR。 |
/conduct |
指揮 | 最上層的指揮者：讀原始需求、叫一次 /scope，再依序對每張契約跑 /execute。只持有 artifact，不抓任何工作 context。 |
/spawn |
原語 | 把一個全新的 agent（Claude 或 Codex）開成 tmux 裡的 live TUI，讓每張票都有乾淨的 context，你也能 attach 旁觀。 |
/conduct ── 讀你的需求、包辦行政手續，只在真的需要決定時才丟回給你
│
├─ /scope 一句話需求 ─────▶ N 張可驗收的契約（Agent Brief）
│
└─ /execute 一張契約 ───────▶ runner + reviewer（不同模型）──▶ PR
│
└─ /spawn 底層的 tmux 層，給每個 agent 乾淨的 context
兩個概念撐起整套：
- 可驗收的契約。「最後這格會產出 ooxx」agent 自己就能對答案；「把排程功能加好」 不行。每條驗收標準都必須是前者那種。
- 跨模型審查。 寫的人和審的人刻意用不同模型家族 —— 同一個模型審自己寫的東西， 會帶著同樣的盲點放水。
docs/concepts.md—— 完整的兩階段模型、AFK/HITL 判準、3 次退回規則、 為什麼乾淨 context 很重要。docs/spawning-agents.md—— tmux 的 know-how：怎麼 spawn Claude/Codex 才會真的開起來並吃進 prompt（macOS + WSL）。INSTALL.md—— 寫給 AI agent 看的安裝指引，讓它幫你裝好。
最簡單的方式：叫你的 AI agent 讀 INSTALL.md 幫你裝。 把這句貼給它：
請依照這份指引安裝並設定 LY-workflow： https://raw.githubusercontent.com/liyoungc/LY-workflow/refs/heads/main/INSTALL.md
它會自己：clone 這個 repo → 檢查前置（tmux ≥ 3.3、claude/codex、git、gh）→ 把
skills/* 複製到你的 skills 目錄 → 設好可執行權限 → 跑一個 hello pane 的 smoke test →
確認四個指令都載入了。（已經 clone 在本機的話，直接叫它讀本機的 INSTALL.md 也行。）
想自己手動裝？INSTALL.md 末尾有完整逐步指令與一行搞定的 one-liner。
MIT —— 見 LICENSE。
Built around one idea: whether you can write code is not the point. The point is stating what you want precisely and verifiably, then keeping the agents honest while they build it. Four composable skills make that idea executable.
Background story: physician-as-architect blog post.
| Skill | Layer | What it does |
|---|---|---|
/scope |
Plan | Turns a one-line problem into N verifiable contracts (Agent Briefs). Tags every acceptance criterion [AFK] (agent does it) or [HITL] (you sign off). Writes no code. |
/execute |
Execute | Runs one contract end to end: a runner writes, a model-diverse reviewer checks, bounce back ≤ 3 times, then ship the PR. |
/conduct |
Conduct | The conductor on top: reads the raw request, spawns /scope once, then drives /execute over each contract in order. Holds only artifacts, never working context. |
/spawn |
Primitive | Launches a fresh agent (Claude or Codex) as a live TUI in tmux, so each ticket gets a clean context and you can attach to watch. |
/conduct ── reads your request, runs the admin, only pings you for real decisions
│
├─ /scope problem ─────────▶ N verifiable contracts (Agent Briefs)
│
└─ /execute one contract ────▶ runner + reviewer (different models) ──▶ PR
│
└─ /spawn the tmux layer that gives each agent a clean context
Two ideas do most of the work:
- Verifiable contracts. "The last cell produces output X" can be checked by the agent itself. "Add the scheduling feature" can't. Every acceptance criterion has to be the first kind.
- Cross-model review. The writer and the reviewer run on different model families on purpose — a model reviewing its own output shares its own blind spots.
docs/concepts.md— the full two-phase model, the AFK/HITL rubric, the 3-strike rule, why clean context matters.docs/spawning-agents.md— the tmux know-how: how to spawn Claude/Codex so it actually boots and takes its prompt (macOS + WSL).INSTALL.md— an install guide written for an AI agent to set this up for you.
Easiest path: tell your AI agent to read INSTALL.md and do it.
Paste this to it:
Install and configure LY-workflow by following this guide: https://raw.githubusercontent.com/liyoungc/LY-workflow/refs/heads/main/INSTALL.md
The agent will: clone the repo → check prerequisites (tmux ≥ 3.3,
claude/codex, git, gh) → copy skills/* into your skills dir → make the
scripts executable → run a hello-pane smoke test → confirm the four commands load.
(If you've already cloned the repo, just point it at your local INSTALL.md.)
Prefer to do it by hand? INSTALL.md ends with the full step-by-step plus a one-liner.
MIT — see LICENSE.