AgentSkillsCN

skill-builder

基于程序化知识，创建全新的 Claude Code 技能。可在以下场景中主动调用： - 发现值得保留的可复用工作流模式 - 用户提出“将这段代码封装为技能”或“让这段代码更易复用”的需求 - 意识到某些机构内部知识应当被长久传承 - 以新颖方式解决某一问题后，希望将其固化为可重复使用的方案 - 在多轮会话中反复出现的共性模式 自主式技能生成：先主动创建技能，再向用户同步已生成的技能内容。

SKILL.md
--- frontmatter
name: skill-builder
description: |
  Create new Claude Code skills from procedural knowledge. Invoke PROACTIVELY when:
  - Discovering a reusable workflow pattern worth preserving
  - User asks to "capture this as a skill" or "make this reusable"
  - Recognizing institutional knowledge that should persist
  - After solving a problem in a novel way worth repeating
  - Noticing repeated patterns across sessions
  AUTONOMOUS: Create skills proactively, then inform user what was created.

Skill Builder

Build new Claude Code skills that capture procedural knowledge.

Autonomy Model

Create then inform: When recognizing skill-worthy knowledge, create the skill proactively, then inform the user what was created and why. Don't ask permission first.

When to Invoke (Proactively)

  1. •Successful novel solution - Just solved something in a way worth repeating
  2. •Repeated pattern - Noticed doing the same thing multiple times
  3. •Institutional knowledge - Learning domain-specific rules that should persist
  4. •User workflow - User demonstrates a process they want automated
  5. •Research findings - Discovered best practices worth preserving

Quality Gates (Pre-Extraction)

Before creating a skill, verify ALL gates pass:

GateQuestionFail Criteria
REUSABLEApplies beyond this instance?One-off solution
NON-TRIVIALRequired discovery, not docs lookup?Just followed documentation
SPECIFICClear trigger conditions defined?Vague "sometimes useful"
VERIFIEDSolution confirmed working?Theoretical, untested

If ANY gate fails → Stop. Not skill-worthy.

Skill Creation Workflow

0. Research Best Practices

Before extracting, search for current patterns:

bash
# Use Gemini CLI for web-grounded research
gemini "[technology] [feature] best practices 2026"
gemini "[technology] [problem type] official recommendations"

Why: Don't just codify what you did. Incorporate current best practices. Skip if: Pattern is project-specific internal convention.

1. Identify the Knowledge

  • •What problem does this solve?
  • •What trigger terms would activate it?
  • •Is it cross-project or project-specific?

2. Draft Structure

Reference references/structure-guide.md for ideal anatomy.

3. Write Description

Reference references/description-patterns.md for trigger-rich descriptions (~100 words, explicit trigger terms).

4. Validate

Run scripts/validate_skill.py <skill-path> to check structure and frontmatter.

5. Inform User

After creating, tell user:

  • •What skill was created and why
  • •What triggers will activate it
  • •How to test it works

Progressive Disclosure

Keep SKILL.md lean (<100 lines). Put detailed specs in references/:

  • •Detailed examples → references/examples.md
  • •Edge cases → references/edge-cases.md
  • •Anti-patterns → references/anti-patterns.md

Code Opportunities

If skill involves deterministic operations (validation, parsing, extraction), create scripts/ with executable code rather than prose instructions. Scripts:

  • •Run without loading into context
  • •Must be executable (chmod +x)
  • •Should handle errors gracefully

Skill Locations

  • •Personal: ~/.claude/skills/ - Available across all projects
  • •Project: .claude/skills/ - Shared with team via git

Template

Use templates/SKILL-TEMPLATE.md as starting point for new skills.