| name | skill-breakdown-workbench |
| description | Analyze Codex Skills and agent manifests. Use when the user wants to拆解 a Skill, explain how a Skill or agent is written, or learn how to write one with bilingual Chinese/English guidance, reusable Skill formulas, and templates, especially when the target is an existing SKILL.md or agents/openai.yaml. |
拆解现有的skill / Existing Skill Breakdown
Overview
This skill turns a target Skill or agent manifest into a bilingual teardown and a writing lesson. It explains what the artifact does, how it is structured, and how to write a stronger version without inventing hidden behavior.
The deliverable is always a Markdown artifact. 每次拆解、写作教学或改进分析都必须生成一个 .md 文件,不能只在对话中返回散文说明。
When to Read References
Read references/guide.md when the user asks for:
- a full Skill-vs-agent comparison
- reusable Skill formulas or authoring formats
- a reusable writing template
- a rewrite or improvement plan
- a stricter checklist for authoring or reviewing a package
Triggers
- “拆解这个 skill 是怎么写的”
- “帮我分析这个 agent manifest,并教我怎么写”
- “Break down this Skill and show me how it is written”
- “Teach me how to write a better Skill and agent from this example”
Workflow
- Identify the target artifact: Skill, agent manifest, or both.
- Read only visible files and treat
SKILL.md, agents/openai.yaml, and linked references as the source of truth.
- Explain the artifact in paired Chinese/English sections:
- 结论 / Verdict
- 结构拆解 / Structure Breakdown
- Skill 写作法 / Skill Writing
- Agent 写作法 / Agent Writing
- 可复用公式 / Reusable Formula
- 可复用模板 / Reusable Template
- Turn the analysis into concrete lessons: triggers, boundaries, workflow, decision rules, references, output contract, and validation.
- When teaching how to write a Skill, include the reusable formula:
Skill = Trigger + Boundary + Workflow + Decision Rules + References + Output Contract + Validation.
- When useful, map the target artifact back into the formula so the user can see how each part is represented in visible files.
- If the user wants a rewrite, keep the original capability and boundary, then produce a draft or patch instead of a generic summary.
- If the task is only review or explanation, do not install, publish, commit, or push anything.
- If the task edits files, verify that every claim and every recommendation maps back to the visible text.
- Write the final teardown to a
.md file and return a clickable link to that file. 将最终拆解写入 .md 文件,并在回复中返回可点击文件链接。
Guardrails
- Do not guess hidden prompts, private APIs, or unpublished files.
- Do not turn a Skill into a persona or soul statement.
- Keep the boundary clear: Skills describe capability and workflow; agents describe entry behavior and invocation details.
- Do not claim installation, publication, or execution unless it actually happened.
- Do not introduce secrets, local absolute paths, or external side effects without explicit permission.
Validation
- For analysis, check that each conclusion is grounded in visible content.
- For authoring, validate
SKILL.md, agents/openai.yaml, and any references before handing off.
- When files change, run the repository's structural checks and a path/secret scan before delivery.
- Confirm the output path ends in
.md, the file exists, and the required bilingual sections are present.
- Do not claim the teardown was delivered until the Markdown file was written successfully.