为什么值得关注

它不是理論藍圖，而是從真實運作的系統中描繪的平面圖——每個範例都來自去識別化的真實事件，並且 v1.1 版本是基於模型實際執行時發現的三個遺漏點修正而來。

适合谁使用

使用 Claude Code 或類似架構的開發者
部署多代理 AI 系統的團隊
希望在模型退役後保留其行為準則的工程師
長期運行 AI 輔助專案的運維人員

典型使用场景

提升中階模型（Opus、Sonnet、Haiku）至 Fable 等級的作業紀律
在專案中標準化代理的派工、升級與驗證流程
從即將停用的頂級模型中擷取並轉移操作知識
透過教訓紀錄（lessons.md）與安全修訂機制維護規則集

项目优势

基於真實跨機器部署驗證，附有去識別化的日期事件範例
專為弱模型設計，不依賴前端模型
對品味判斷與模糊問題提供明確的退場路徑
內建演化機制：教訓先寫入 lessons.md，再按維護規則收編

使用前须知

不解決品味判斷與問題定義模糊的難題，需要誠實升級或求助
環境特定資訊（模型 ID、工具名稱）每部署一次都必須重新驗證
初始需要投入時間將通用規則適應到具體專案環境

README 快速开始

project-ops-template

Keep the Fable state after Fable is gone.

Claude Fable 5 — the top judgment tier — sunsets for subscription use on 2026-07-07. The point of this repo is to capture that judgment as portable doctrine: a rules layer that guides a mid-tier model (Opus, Sonnet-class, even Haiku-class) into operating at the Fable level — knowing when to delegate, when to escalate, when to stop and ask, how to verify without lying to itself, and how the rules themselves get safely revised over time.

You lose the model. You keep the operating discipline that made it good. That transfer — top-tier judgment → concrete, executable rules a lesser model can run — is the whole idea, and it's why the rules are written concrete-to-the-point-of-executable, every one with a decision test and worked examples.

Works with Claude Code, or any harness built around a CLAUDE.md-style entry file plus a subagent/dispatch mechanism.

Why this one, and not a prompt-generated template

This is not a set of rules an LLM was asked to draft in one pass and never ran. It is the operating manual of an actual long-running multi-agent deployment, extracted, sanitized, and open-sourced.

The concrete difference: after the first full version was written, the strongest model available in that deployment was handed the eight-step command loop and told to run it, for real, against a real task — not to review it, to use it. It came back and, in its own words, patched three concrete gaps the desk-review pass had missed (a role-assignment ambiguity, a missing "worker lane" for delegated subtasks, and a fuzzy test for telling a genuine limitation from an excuse). That's the v1.1 revision baked into these files. The rules have also since been the actual dispatch backbone for a real cross-machine deployment and ongoing multi-agent collaboration — every worked example in ops/00-command-loop.md and ops/50-diagnosis.md is a dated, real incident with the identifying details removed, not a hypothetical.

Put simply: most templates are a blueprint. This one is a floor plan traced from a house people actually live in — the floorboards that creak are marked, because someone found them by stepping on them.

What's in the box

project-ops-template/
├── README.md                  ← this file (safe to delete after deployment)
├── CLAUDE.md.template         ← a lean, routing-only CLA

项目描述

Battle-tested model-dispatch doctrine for Claude Code — portable ops rules that let Sonnet/Haiku reliably run projects. Iterated in real deployments, not one-shot generated.

claude-project-ops