skill-factory

多代理管道编排器，从头开始构建新技能。在以下情况下使用：(1) 您有一个技能想法，并希望将其从概念转化为可投入市场的包，(2) 您需要一个结构化的 7 阶段流程：市场研究 → 规划 → 架构 → 构建 → 审计 → 文档 → 定价，(3) 您希望每个阶段作为独立的、顺序的代理调用运行，没有嵌套会话。触发：“创建新技能”、“构建技能”、“技能工厂”、“新技能管道”。

# Skill Factory

A 7-stage multi-agent pipeline that takes a raw skill idea and produces a complete, market-ready skill package. Each stage is an isolated CLI agent call — no nested sessions, no parallelism.

## Overview

```
idea → market → planner → arch → builder → auditor → docs → pricer → done
```

All state is written to disk between stages. Any stage can be re-run independently.

## Workflow

### Step 1: Initialize a skill idea

```bash
python3 {baseDir}/scripts/init_pipeline.py "My Skill Idea" --workspace /tmp/sf-my-skill
```

This creates a workspace directory with `idea.md` pre-filled. Review and edit `idea.md` before running the pipeline.

### Step 2: Run the full pipeline

```bash
bash {baseDir}/scripts/pipeline.sh --workspace /tmp/sf-my-skill
```

The pipeline runs all 7 agents sequentially. Each agent reads previous outputs and writes its own report into the workspace.

### Step 3: Collect outputs

After the pipeline completes, the workspace contains:

```
workspace/
├── idea.md          # Your original idea (input)
├── market.md        # Market research report
├── plan.md          # Product plan with requirements
├── arch.md          # Architecture and file structure
├── skill/           # The built skill (SKILL.md + scripts/ + references/)
├── audit.md         # Quality audit report
├── docs_review.md   # Documentation review
└── pricing.md       # Pricing and positioning
```

### Step 4: Review gates

Each stage has a gate. If a gate fails, the pipeline stops with a non-zero exit code and reports which stage failed. Fix the issue and re-run from that stage:

```bash
bash {baseDir}/scripts/pipeline.sh --workspace /tmp/sf-my-skill --from auditor
```

## Stages at a glance

| Stage    | Agent ID   | Input              | Output         |
|----------|------------|--------------------|----------------|
| market   | market     | idea.md            | market.md      |
| planner  | planner    | idea.md, market.md | plan.md        |
| arch     | arch       | plan.md            | arch.md        |
| builder  | builder    | arch.md, plan.md   | skill/         |
| auditor  | auditor    | skill/             | audit.md       |
| docs     | docs       | skill/, audit.md   | docs_review.md |
| pricer   | pricer     | all outputs        | pricing.md     |

Read [references/agents.md](references/agents.md) for full agent descriptions, models, and prompt guidance.

Read [references/pipeline.md](references/pipeline.md) for detailed gate definitions, failure modes, and re-run strategies.

## Key Constraints

- **No nested sessions** — every agent runs as a top-level isolated CLI call
- **No parallelism** — stages run strictly sequentially; each depends on the previous
- **File-based state** — all inter-stage communication goes through the workspace filesystem
- **Idempotent stages** — re-running a stage overwrites its output without affecting others

## When NOT to Use This Skill

- You only need a quick, simple skill with no research phase — use `init_skill.py` directly
- You are iterating on an existing skill — edit files directly
- You want parallel research across multiple skill ideas — run multiple pipeline instances in separate workspaces

---

## 中文说明

# Skill Factory

一个 7 阶段的多代理管道，将原始的技能想法转化为完整、可投入市场的技能包。每个阶段都是一个独立的 CLI 代理调用——没有嵌套会话，没有并行。

## 概述

```
idea → market → planner → arch → builder → auditor → docs → pricer → done
```

所有状态在各阶段之间都会写入磁盘。任何阶段都可以独立地重新运行。

## 工作流程

### 步骤 1：初始化技能想法

```bash
python3 {baseDir}/scripts/init_pipeline.py "My Skill Idea" --workspace /tmp/sf-my-skill
```

这会创建一个工作区目录，其中预填充了 `idea.md`。在运行管道之前，请检查并编辑 `idea.md`。

### 步骤 2：运行完整管道

```bash
bash {baseDir}/scripts/pipeline.sh --workspace /tmp/sf-my-skill
```

管道会依次运行所有 7 个代理。每个代理读取先前的输出，并将自己的报告写入工作区。

### 步骤 3：收集输出

管道完成后，工作区包含：

```
workspace/
├── idea.md          # Your original idea (input)
├── market.md        # Market research report
├── plan.md          # Product plan with requirements
├── arch.md          # Architecture and file structure
├── skill/           # The built skill (SKILL.md + scripts/ + references/)
├── audit.md         # Quality audit report
├── docs_review.md   # Documentation review
└── pricing.md       # Pricing and positioning
```

### 步骤 4：审查关卡

每个阶段都有一个关卡 (gate)。如果某个关卡失败，管道会以非零退出码停止，并报告失败的阶段。修复问题后从该阶段重新运行：

```bash
bash {baseDir}/scripts/pipeline.sh --workspace /tmp/sf-my-skill --from auditor
```

## 各阶段一览

| 阶段     | 代理 ID    | 输入               | 输出           |
|----------|------------|--------------------|----------------|
| market   | market     | idea.md            | market.md      |
| planner  | planner    | idea.md, market.md | plan.md        |
| arch     | arch       | plan.md            | arch.md        |
| builder  | builder    | arch.md, plan.md   | skill/         |
| auditor  | auditor    | skill/             | audit.md       |
| docs     | docs       | skill/, audit.md   | docs_review.md |
| pricer   | pricer     | all outputs        | pricing.md     |

阅读 [references/agents.md](references/agents.md) 获取完整的代理描述、模型和提示指南。

阅读 [references/pipeline.md](references/pipeline.md) 获取详细的关卡定义、失败模式和重新运行策略。

## 关键约束

- **无嵌套会话**——每个代理都作为顶层独立 CLI 调用运行
- **无并行**——各阶段严格按顺序运行；每个阶段都依赖于前一个
- **基于文件的状态**——所有阶段间通信都通过工作区文件系统进行
- **幂等阶段**——重新运行某个阶段会覆盖其输出，而不影响其他阶段

## 何时不应使用此技能

- 你只需要一个快速、简单、无需研究阶段的技能——直接使用 `init_skill.py`
- 你正在迭代一个现有的技能——直接编辑文件
- 你想对多个技能想法进行并行研究——在单独的工作区中运行多个管道实例