Compound Learning Loop. 跨 sprint 积累的经验、模式、约束沉淀到 lessons.md, 后续 sprint 自动加载.
Scanned 9/5/2026
Install to Claude Code
npx -y skills add WenJunDuan/Rlues --skill compound --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Compound?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/wenjunduan-compound-ac9e9821)More formats (shields.io, HTML) on the badges page.
---
name: compound
description: |
Compound Learning Loop. 跨 sprint 积累的经验、模式、约束沉淀到 lessons.md, 后续 sprint 自动加载.
effort: medium
---
# /compound — 经验积累循环 (v9.6.2)
## 概念
Compound Learning Loop = 每次 polish / review 完成后, 把"做对的 pattern + 踩的坑 + 硬约束"沉淀到 `.ai_state/lessons.md`. 下一个 sprint 进入 plan stage 时, 主 agent **必须** read lessons.md 一次, 防止重复犯错.
## 写入时机
| 时机 | 触发 | 写入什么 |
|---|---|---|
| review_pass1 完成 | evaluator 写 VERDICT 时 | 1 行 pattern (做对了什么) |
| polish 完成 | cleanup-pass.md 写完时 | 1 行 pitfall (踩了什么坑) |
| 用户显式 `/compound add` | 用户主动 | 任意经验 |
| reviewer 发现新 P0 finding | findings 写完时 | constraint (硬约束) |
## 格式
```markdown
# Project Lessons
## [2026-05-23 sprint-2] Refactor JWT module
- Pattern: 用 RS256 instead of HS256 for stateless verification across services [executed]
- Pitfall: useEffect deps 漏掉 user.id 导致重渲染 inf loop [inspected]
- Constraint: 任何 JWT secret 改动必须更新所有 microservice 同时部署 [executed]
## [2026-05-22 sprint-1] Initial setup
- Pattern: TDD 严格执行后, refactor 阶段省了 40% 时间 [executed]
```
## 读取时机
- **plan stage 开始**: 主 agent Read `.ai_state/lessons.md` (强制, 不允许跳)
- **review stage 开始**: reviewer 读 lessons.md 看是否本 sprint 触犯了已知 pitfall
## 增量规则
- lessons.md ≤ 200 行: 持续追加
- lessons.md > 200 行: 主 agent 在 polish 后做"压缩 pass", 把相似 pattern 合并, 保留高密度
- lessons.md > 500 行: 拆分 archive (`lessons-2026-Q1.md`), 主 lessons.md 只留最近 3 个 sprint
## 约束
- 每条 lesson 必须带状态标 (`[executed]` / `[inspected]` / `[deferred]`)
- 不允许长篇大论 (一条 ≤ 30 字)
- 不允许重复 pattern (相同 pattern 第二次出现 → 升级为铁律候选)
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!