<h1 align="center">Cleaver</h1>
<p align="center"><strong>Take any product. Cleave it into the prompts that built it.</strong></p>
<p align="center">
A <a href="https://docs.anthropic.com/en/docs/claude-code/skills">Claude Code skill</a> that reverse-engineers finished products into buildable prompts.
</p>
<p align="center">
<a href="./README.zh-CN.md">中文</a> · <strong>English</strong>
</p>
<p align="center">
<img src="https://img.shields.io/badge/scenarios-17-green" />
<img src="https://img.shields.io/badge/quality_dimensions-12-blue" />
<img src="https://img.shields.io/badge/pass_rate-85%25-brightgreen" />
<img src="https://img.shields.io/badge/without-32%25-lightgrey" />
<img src="https://img.shields.io/badge/license-MIT-yellow" />
</p>
---
> **Cleaver doesn't read minds.** It reads products.
> Observable decisions, explicit assumptions, rebuildable prompts.
> No leaked internals, no claimed secrets.
## The problem with "make it like X"
```
❌ Make a dashboard like Linear.
→ Copies the chrome, misses the engine.
→ No speed philosophy, no keyboard flow, no soul.
→ Ship something that looks like Linear but feels like Jira.
✅ Build an issue tracker whose promise is "nothing slows you down".
Keyboard-first inbox, instant command palette, no modal editing.
Done when you can triage 50 issues without reaching for the mouse.
Not yet: roadmap, docs, chat, analytics.
```
Most vibe coding fails at the prompt, not the code.
Cleaver studies products that already work — and extracts the prompts that would have built them.
---
## What it cleaves
Give it **anything finished** — it hands you the prompts to rebuild it.
| Input | Output |
|-------|--------|
| Screenshot | Layer-by-layer visual deconstruction → build prompts |
| URL | Live product analysis → scoped rebuild prompts |
| Code repo | Architecture extraction → spec + build chain |
| Verbal description | Product archetype inference → prototype prompts |
| A single feature | Trigger → change → transition → prompt |
| A physical object | Sensory + interaction profile → experience spec |
**Output modes** — prompts for vibe coding, PRDs, design briefs, service blueprints, or guided learning.
---
## 17 scenarios. 85% pass rate.
<img src="./docs/benchmark.svg" alt="benchmark comparison chart" width="100%" />
| Metric | With Cleaver | Without | Delta |
|--------|-------------|---------|-------|
| Average pass rate | **85%** | 32% | **+53pp** |
| Soul capture | 17/17 (100%) | 7/17 (41%) | +59pp |
| Scope control | 17/17 (100%) | 6/17 (35%) | +65pp |
| Teaching annotations | 16/17 (94%) | 0/17 (0%) | +94pp |
<img src="./docs/dimensions.svg" alt="dimension coverage chart" width="100%" />
> Each scenario graded on 12 dimensions: product analysis, prompt quality, scope control, build order, domain framework, teaching value, and more. Full rubric in [`evals/rubric.md`](evals/rubric.md).
---
## Four paths. Pick your depth.
| Path | Prompts | Time | When |
|------|---------|------|------|
| **Minimal** | 1 (2-3 sentences) | instant | "Just the soul" |
| **Fast Track** | 2-3 | ~30 min | "Ship something tonight" |
| **Standard Build** | 5-8 | hours | "Rebuild the whole thing" |
| **Learning Deep-Dive** | 5-10 (annotated) | hours | "Teach me to think in prompts" |
Every path (except Minimal) starts with **Prompt 0** — a foundation prompt that establishes project DNA (stack, structure, conventions, done condition) so every subsequent prompt builds instead of re-establishing context.
---
## Ten domains. Ten frameworks.
| | | | | |
|---|---|---|---|---|
| Web App / SaaS | Mobile App | Landing Page | Animation | CLI Tool |
| Design System | Game | API / Backend | AI Product | Service / Physical |
Games get MDA analysis. APIs get contract-driven decomposition. AI products get system prompt architecture. Every domain has its own lens — one framework doesn't fit all.
---
## Install
```bash
npx skills add taekchef/cleaver
```
Then just describe what you want to deconstruct:
```
> 拆解 Stripe 的 API 设计理念
> Deconstruct Figma — I want to build something similar in 30 minutes
> 帮我用最少的话拆解 Notion
> Break down the iOS delete-app wiggle animation
```
---
## Examples
### Minimal — Notion in 3 sentences
```
做一个"万物皆 block"的工作空间：每一段文字、每一张图、每一行数据库都是同一颗原子积木，
可以嵌套、拖拽、变形、关联——像乐高一样拼出笔记、文档、看板、日历、Wiki 任何形态。
打开是一张白纸，干净到没有存在感，但底层是一个图结构的数据库引擎，
让个人和团队在同一块画布上实时协作、自定义任何工作流。
不要做固定模板的 SaaS，要做用户自己造工具的平台——Notion 卖的不是功能，是"你可以自己搭"的创造力。
```
### Fast Track — Wordle (3 prompts, MDA framework)
Soul: "one sentence explains the rules". Foundation (grid + keyboard) → Game logic (with duplicate-letter edge cases) → Animation + Share (the viral engine).
→ [Full output](examples/wordle-game.md)
### Standard Build — Stripe API (6 prompts)
Philosophy → Data model → API surface (CRUD, cursor pagination, expand) → Operational contracts (idempotency, webhook signatures) → Error model (three-layer classification) → Developer experience.
→ [Full output](examples/stripe-api.md)
### Learning Deep-Dive — Perplexity (6 annotated prompts)
AI product deconstruction with system prompt architecture. Soul: "every answer has evidence". Each prompt comes with a "why this works" annotation.
→ [Full output](examples/perplexity-ai-product.md)
### Fast Track — "Tinder for restaurants" (verbal-only, 3 prompts)
User says one sentence — Cleaver infers the product archetype, identifies "decision fatigue killer", and builds.
→ [Full output](examples/tinder-restaurant.md)
---
## How it works
```
Anything ──► Read the product ──► Cleave into layers ──► Write prompts
finished (observe + infer) (6-layer framework, (12 prompt patterns,
domain-specific) path-specific gate)
```
**12 prompt patterns** across three categories:
| Build prompts | Product docs | Technical contracts |
|---|---|---|
| Intent-first | PRD generator | System prompt |
| Spec-driven | Design brief | API contract |
| Iterative chain | Experience-to-Spec | |
| Not-to-dos | GDD generator | |
| Example-driven | | |
| Test-first | | |
→ [`references/patterns/build-prompts.md`](references/patterns/build-prompts.md) · [`product-docs.md`](references/patterns/product-docs.md) · [`technical-contracts.md`](references/patterns/technical-contracts.md)
---
## Architecture
```
cleaver/
├── SKILL.md # The skill itself (~195 lines)
├── evals/
│ ├── rubric.md # 12-dimension grading criteria
│ ├── benchmark.json # Aggregated results
│ └── build_benchmark.py # Eval → benchmark pipeline
├── docs/
│ ├── benchmark.svg # Scenario comparison chart
│ ├── dimensions.svg # Dimension coverage chart
│ └── generate_charts.py # Benchmark → SVG pipeline
├── references/
│ ├── domains/ # 10 domain-specific strategies
│ └── patterns/ # 12 prompt pattern references
└── examples/ # Real teardown outputs
```
---
## Responsible use
For learning, inspiration, and legitimate remixing.
Not for copying proprietary assets, impersonating brands, or bypassing access controls.
**Preserve the lesson, not the identity.**
Extract patterns and principles — avoid copying names, branding, or proprietary implementation.
---
## License
MIT