claude-code-dev-workflow
Stop vibe-coding from vague prompts. A set of Claude Code plugins that give the agent an opinionated spec-to-ship workflow: it grills your half-baked idea into a real PRD, then drives it through plan → TDD → review → summary.
English | 中文
Why
Most bad AI output isn't a coding problem — it's a spec problem. You hand the agent a fuzzy sentence, it happily builds the wrong thing, and you burn an hour finding out. These plugins put a process in front of the code:
- forge-prd interrogates your idea before a line is written — like a red team trying to find the hole in your requirement.
- dev-flow then carries that spec through a disciplined 8-step pipeline so nothing gets skipped.
- dev-flow-lite is the 4-step version for when you just need to move.
Quickstart
/plugin marketplace add Today-Hbw/claude-code-dev-workflow
/plugin install forge-prd@claude-code-dev-workflow
/plugin install dev-flow@claude-code-dev-workflow
/plugin install dev-flow-lite@claude-code-dev-workflow
Then:
/forge-prd <a rough idea, a screenshot, a PDF — anything>
/dev-flow <path to the PRD you just forged>
The plugins
| Plugin | What it does |
|---|---|
| ⚒️ forge-prd | Forge — don't polish. Takes any raw material (a rough thought, image, PDF, HTML, MD) and grills it in three escalating layers until a clear, build-ready PRD falls out. |
| 🔁 dev-flow | The full 8-step spec-to-delivery pipeline: read PRD → grill → Q&A record → tech plan → code (TDD) → self-test → code review → summary. Resumable, config-driven. |
| ⚡ dev-flow-lite | The 4-step version: read PRD → align → code → review. No docs, no resume — just momentum for small changes. |
⚒️ Spotlight: forge-prd
Forge, not polish. The requirement itself might be wrong — so the plugin attacks it like a red team before you build.
Three escalating layers of interrogation:
- Absorb — reads all your material (MD / PDF / HTML / images); interprets images and asks you to confirm.
- Confirm understanding — plays back "what I think you want is X" before going deeper.
- Three-layer grilling — ambiguity → boundary → assumption, each layer gated by your confirmation.
- Output — a structured
<name>.forged.mdyou can hand straight to development.
🔁 Spotlight: dev-flow
An opinionated 8-step workflow Claude follows automatically, emitting standard docs at each stage:
| # | Step | Output | Key method |
|---|---|---|---|
| 1 | Read PRD | requirement summary | scan + extract |
| 2 | Grill | questions + glossary | decision-tree traversal |
| 3 | Q&A record | 问答记录.md | structured Q/A |
| 4 | Tech plan | <task>/计划.md | agent-brief style |
| 5 | Code | source files | vertical slice + TDD |
| 6 | Self-test | test report | behavior tests + checks |
| 7 | Code review | review notes | verify-first (run tests before reviewing) |
| 8 | Summary | 总结.md | deliverable record |
Control flags: --skip 5,6, --from 4, --only 3, --resume, --overview. See plugins/dev-flow.
Star this repo ⭐
If this saved you an hour of rework, a star helps other Claude Code users find it.
Credits
Design inspired by mattpocock/skills and anthropics/claude-plugins-official.
License
MIT
中文
别再用一句模糊的话去 vibe coding。 一组 Claude Code 插件,给 agent 装上一套有主见的**"从需求到交付"工作流**:先把你半生不熟的想法拷问成一份真正的 PRD,再驱动它走完 规划 → TDD → 评审 → 总结。
English | 中文
为什么
大多数 AI 产出的垃圾不是编码问题,而是需求问题。你丢给它一句模糊的话,它高高兴兴地把错的东西做出来,你花一小时才发现。这组插件在写代码之前先加一道流程:
- forge-prd 在你动笔前就质疑你的想法——像红队一样专找需求里的漏洞。
- dev-flow 再带着这份需求走完一条纪律严明的 8 步流水线,一步不落。
- dev-flow-lite 是 4 步简版,只想快速推进时用它。
快速开始
/plugin marketplace add Today-Hbw/claude-code-dev-workflow
/plugin install forge-prd@claude-code-dev-workflow
/plugin install dev-flow@claude-code-dev-workflow
/plugin install dev-flow-lite@claude-code-dev-workflow
然后:
/forge-prd <一个粗略想法、一张截图、一个 PDF——什么都行>
/dev-flow <你刚锻造出的 PRD 路径>
插件列表
| 插件 | 说明 |
|---|---|
| ⚒️ forge-prd | 锻造,不是打磨。把任何原料(粗略想法、图片、PDF、HTML、MD)分三层递进强势拷问,直到逼出一份清晰、可直接开工的 PRD。 |
| 🔁 dev-flow | 完整 8 步从需求到交付流水线:读 PRD → 拷问 → 问答记录 → 技术方案 → 编码(TDD) → 自测 → CR → 总结。可断点续传、配置驱动。 |
| ⚡ dev-flow-lite | 4 步简版:读 PRD → 对齐 → 编码 → 评审。不写文档、不续传,小改动直接冲。 |
⚒️ 重点:forge-prd
锻造,而非打磨。 需求本身可能就是错的——所以插件在你动手前像红队一样攻击它。
四个阶段:
- 吸收 — 读取所有素材(MD / PDF / HTML / 图片),主动解读图片并请你确认。
- 理解确认 — 先复述"我理解你想要的是 X",大方向对了再深入。
- 三层拷问 — 模糊层 → 边界层 → 假设层,每层结束由你确认才进入下一层。
- 输出 — 一份结构化的
<原名>.forged.md,可直接交给研发。
🔁 重点:dev-flow
Claude 自动遵循的 8 步流程,每一步产出标准文档:
| # | 步骤 | 输出 | 关键方法 |
|---|---|---|---|
| 1 | 读 PRD | 需求摘要 | 扫描 + 提取 |
| 2 | 拷问 | 提问 + 术语表 | 决策树遍历 |
| 3 | 问答记录 | 问答记录.md | 结构化 Q/A |
| 4 | 技术方案 | <任务>/计划.md | Agent Brief 风格 |
| 5 | 编码 | 代码文件 | 垂直切片 + TDD |
| 6 | 自测 | 测试报告 | 行为测试 + 检查 |
| 7 | CR | 审查意见 | 验证前置(先跑测试再审查) |
| 8 | 总结 | 总结.md | 交付物记录 |
控制参数:--skip 5,6、--from 4、--only 3、--resume、--overview。详见 plugins/dev-flow。
点个 Star ⭐
如果它帮你省下了一小时返工,一个 star 能让更多 Claude Code 用户找到它。
致谢
设计参考 mattpocock/skills 与 anthropics/claude-plugins-official。
License
MIT