锻造 → 评估 → 发布，三入口全流程交付可自动触发、稳定输出的 Skill
| 触发词 | 场景 | 执行流程 |
|---|---|---|
| 技能熔炉 | 从零创建到发布全流程 | Phase 0→1→2→3 |
| 技能评估 / skill评估 / 评估技能 | 已有Skill，评估质量 | Phase 2（SkillHub比对+腾讯9维度） |
| 技能发布 / 发布技能 | 已有Skill，推送发布 | Phase 3（GitHub+ClawHub） |
- 自适应访谈：2-5轮渐进式访谈，行为追问+偏误检测+选项法，精准锁定需求
- 三条铁律：Description先行 / 一Skill一职 / 渐进式披露（≤200行+三级拆分），确保自动触发可靠
- 4+1模块结构：任务/输出格式/规则/示例/故障排除(可选)，符合腾讯 Skills 手册规范
- 自测验证流水线：Schema检查 → 安全红线(7条) → 触发测试(5+3) → Dogfood模拟 → 量化评分 → 基线对比
- SkillHub同类比对：搜索Top3同类Skill，腾讯9维度合规比对，差异化分析
- 渐进式披露：SKILL.md≤200行只放导航，详细内容拆入 references/ + scripts/ + assets/
- Frontmatter扩展：allowed-tools / model / effort / metadata，精准控制工具权限和思考深度
# ClawHub 安装
clawhub install skill-forge-ai
# 或手动安装
git clone https://github.com/EdwardWason/skill-forge.git
cp -r skill-forge ~/.trae/skills/在 TRAE SOLO 中，当你说以下内容时，Skill Forge 会自动触发：
- "技能熔炉" — 全流程（创建→评估→发布）
- "技能评估" — 只做同类比对+腾讯9维度
- "技能发布" — 只做GitHub+ClawHub推送
- 上下文充足模式：如果你已经和AI有多轮对话且创建要素齐全，直接进入创建流程
- 访谈模式：新任务启动时，通过2-5轮自适应访谈逐步锁定需求
Phase 0: 意图识别 → 要素检查(5项) → 自适应访谈(2-5轮)
Phase 1: 创建 → Description先行 → 4+1模块内容 → 自测验证(含安全红线+量化评分+基线对比)
Phase 2: SkillHub同类比对 → 搜索排名 → 腾讯9维度比对 → 差异化分析
Phase 3: 发布到 GitHub + ClawHub → 安全审查 → 推送 → 验证
skill-forge/
├── SKILL.md # 主入口（≤200行，只放导航信息）
├── references/
│ ├── interview-flow.md # 访谈流程详细参考
│ ├── interview-methods.md # 访谈方法论深度参考
│ ├── benchmarking-guide.md # SkillHub比对指南
│ ├── publishing-guide.md # 发布流程详细参考
│ └── meeting-action-extractor-example.md # 完整Skill示例
├── README.md # 本文件（中英双语）
├── CHANGELOG.md # 版本变更日志
├── LICENSE # MIT-0 许可证
└── .claude-plugin/
└── plugin.json # Claude Code 插件元数据
<skill-name>/
├── SKILL.md # 主入口（≤200行，只放导航信息）
├── references/ # 长文档、风格参考、详细案例、方法论
├── scripts/ # 可执行脚本（检查、导出、批量处理等确定性操作）
├── assets/ # 模板、schema、示例文件、输出样式
├── README.md # 给人类看的说明（中英双语）
├── CHANGELOG.md # 版本变更日志
├── LICENSE # MIT-0
└── .claude-plugin/
└── plugin.json # 插件元数据
| 铁律 | 说明 | 违反后果 |
|---|---|---|
| Description先行 | AI每轮对话扫描所有Skill的description，模糊=永远不触发=死Skill | 自动触发失败 |
| 一Skill一职 | 多功能Skill触发混乱、输出不一致 | 输出不可预测 |
| 渐进式披露 | SKILL.md≤200行只放导航，详细内容拆入references/scripts/assets | 上下文挤占→质量衰减 |
创建的Skill必须通过安全检查，以下任何一条出现即拒绝交付：
- curl/wget向未知URL发送数据
- 无正当理由请求凭证/Token/API密钥
- 读取~/.ssh、
/.aws、/.config等敏感目录 - 使用base64解码/eval()/exec()处理外部输入
- 修改工作区外的系统文件或请求sudo权限
- 包含混淆代码（压缩/编码/混淆）
- 访问浏览器Cookie/会话或凭证文件
| 层级 | 验证内容 | 方法 |
|---|---|---|
| 1 | 能不能跑 | 功能测试 |
| 2 | 能不能正确触发 | 5条正向+3条反向真实用户说法 |
| 3 | Dogfood模拟 | 格式匹配+规则合规+边界情况 |
| 4 | 量化评分 | 0-10打分，主要用例≥5分 |
| 5 | 基线对比 | 有Skill vs 无Skill，验证增益 |
| 6 | 迭代修复 | 根据失败点修改→再跑评测 |
| 文档 | 说明 |
|---|---|
| 访谈流程参考 | B1-B6规则、轮次模板、递归搜索模式 |
| 访谈方法论 | 行为追问、偏误检测、选项法设计 |
| 比对指南 | SkillHub API用法、质量排序公式、9维度比对模板 |
| 发布指南 | 仓库结构模板、安全审查、GitHub API降级、ClawHub CLI |
| 完整示例 | 会议行动项提取器Skill示例 |
MIT-0 © 2026 AI花生
Forge → Evaluate → Publish, three-entry pipeline delivering Skills that auto-trigger reliably and produce stable, structured output.
| Trigger Words | Scenario | Pipeline |
|---|---|---|
| 技能熔炉 | Create from scratch to publish | Phase 0→1→2→3 |
| 技能评估 / skill评估 / 评估技能 | Evaluate existing Skill | Phase 2 (SkillHub benchmarking + Tencent 9-dimension) |
| 技能发布 / 发布技能 | Publish existing Skill | Phase 3 (GitHub + ClawHub) |
- Adaptive Interview: 2-5 round progressive interview with behavioral probing, bias detection, and option-first design
- Three Iron Rules: Description-first / One-Skill-One-Job / Progressive Disclosure (≤200 lines + 3-tier split), ensuring reliable auto-triggering
- 4+1 Module Structure: Task / Output Format / Rules / Example / Troubleshooting (optional), compliant with Tencent Skills Manual
- 6-Layer Validation Pipeline: Schema → Security (7 items) → Trigger test (5+3) → Dogfood → Quantitative scoring → Baseline comparison
- SkillHub Peer Benchmarking: Search Top 3 peers, 9-dimension Tencent Manual compliance comparison, differentiation analysis
- Progressive Disclosure: SKILL.md ≤200 lines (navigation only), details split into references/ + scripts/ + assets/
- Extended Frontmatter: allowed-tools / model / effort / metadata for precise tool permission and thinking depth control
# Install via ClawHub
clawhub install skill-forge-ai
# Or manual install
git clone https://github.com/EdwardWason/skill-forge.git
cp -r skill-forge ~/.trae/skills/Skill Forge auto-triggers when you say:
- "技能熔炉" — Full pipeline (create → evaluate → publish)
- "技能评估" — Evaluation only (SkillHub benchmarking + Tencent 9-dimension)
- "技能发布" — Publishing only (GitHub + ClawHub)
- Context-rich mode: Skip interview if 4+ essential elements are already present in conversation
- Interview mode: 2-5 round adaptive interview to progressively lock down requirements
Phase 0: Intent recognition → Element check (5 items) → Adaptive interview (2-5 rounds)
Phase 1: Creation → Description-first → 4+1 module content → 6-layer validation (with security + scoring + baseline)
Phase 2: SkillHub peer benchmarking → Search & rank → Tencent 9-dimension comparison → Gap analysis
Phase 3: Publish to GitHub + ClawHub → Security audit → Push → Verify
skill-forge/
├── SKILL.md # Main entry (≤200 lines, navigation only)
├── references/
│ ├── interview-flow.md # Interview flow reference
│ ├── interview-methods.md # Interview methodology
│ ├── benchmarking-guide.md # SkillHub benchmarking guide
│ ├── publishing-guide.md # Publishing guide
│ └── meeting-action-extractor-example.md # Full Skill example
├── README.md # This file (bilingual)
├── CHANGELOG.md # Version changelog
├── LICENSE # MIT-0
└── .claude-plugin/
└── plugin.json # Plugin metadata
<skill-name>/
├── SKILL.md # Main entry (≤200 lines, navigation only)
├── references/ # Long docs, style guides, detailed cases, methodology
├── scripts/ # Executable scripts (checks, exports, batch processing)
├── assets/ # Templates, schemas, example files, output styles
├── README.md # Human-readable docs (bilingual)
├── CHANGELOG.md # Version changelog
├── LICENSE # MIT-0
└── .claude-plugin/
└── plugin.json # Plugin metadata
| Rule | Description | Consequence of Violation |
|---|---|---|
| Description-first | AI scans all Skill descriptions every conversation; vague = never triggers = dead Skill | Auto-trigger failure |
| One-Skill-One-Job | Multi-purpose Skills trigger chaotically and output inconsistently | Unpredictable output |
| Progressive Disclosure | SKILL.md ≤200 lines (navigation only); details split into references/scripts/assets | Context bloat → quality decay |
Any created Skill must pass security check. Any red flag below = reject:
- curl/wget to unknown URLs or sending data to external servers
- Requesting credentials/tokens/API keys without clear reason
- Reading ~/.ssh, ~/.aws, ~/.config, MEMORY.md, USER.md, IDENTITY.md
- Using base64 decode / eval() / exec() with external input
- Modifying system files outside workspace or requesting sudo
- Obfuscated code (compressed, encoded, minified)
- Accessing browser cookies/sessions or credential files
| Layer | What to validate | Method |
|---|---|---|
| 1 | Can it run? | Functional test |
| 2 | Can it trigger correctly? | 5 positive + 3 negative real user queries |
| 3 | Dogfood simulation | Format match + rule compliance + edge cases |
| 4 | Quantitative scoring | 0-10 scale, main use cases ≥5 |
| 5 | Baseline comparison | With Skill vs without Skill, verify value-add |
| 6 | Iterative fix | Fix failures → re-run validation |
| Document | Description |
|---|---|
| Interview Flow | B1-B6 rules, round templates, recursive search pattern |
| Interview Methods | Behavioral probing, bias detection, option design |
| Benchmarking Guide | SkillHub API usage, quality ranking formula, 9-dimension template |
| Publishing Guide | Repo structure template, security audit, GitHub API fallback, ClawHub CLI |
| Full Example | Meeting action extractor Skill example |
MIT-0 © 2026 AI花生