create-skill（创造技能）

Create Skill

Research-first pipeline for producing Agent Skills that meet the Agent Skills Specification. Every created skill is backed by cited documentation, includes workflow templates for procedural tasks, and never lets the agent substitute patterns for real docs.

Context Loading

Load references based on what the current phase needs. Do NOT load everything upfront.

Skip guidance based on skill type:

Orchestration Flow

Follow these phases in order. Each phase depends on the previous one's output. Track progress through each phase using available task tracking tools — this is a 9-phase process and always qualifies for task tracking.

Phase 0: Load Skill-Development Guidance

Ground the pipeline in official skill development best practices:

1. Search for a local skill-development plugin: look for skill-development/SKILL.md under ~/.claude/plugins/
2. If found locally: invoke Skill("plugin-dev:skill-development") and read its references
3. If not found locally: fetch from the official plugins marketplace:

- https://raw.githubusercontent.com/anthropics/claude-plugins-official/main/plugins/plugin-dev/skills/skill-development/SKILL.md - https://raw.githubusercontent.com/anthropics/claude-plugins-official/main/plugins/plugin-dev/skills/skill-development/references/skill-creator-original.md

1. Keep this guidance in context — the authoring phase needs it

If fetch fails: proceed without it. The quality-criteria.md and workflow-author.md contain enough guidance. Note the gap in the requirements document.

Phase 1: Bootstrap

bash skills/create-skill/scripts/ensure_spec_repo.sh

This ensures ~/.agent-skills-spec contains the latest Agent Skills Specification.

If clone/pull fails: check network connectivity. If offline, check if ~/.agent-skills-spec already exists from a prior run. If it does, proceed with stale data and note it. If it doesn't exist at all, warn the user — spec compliance cannot be verified without it.

Phase 2: Install Skill-Creator

Check if ~/.claude/skills/skill-creator/SKILL.md exists. If not:

claude install-skill github:anthropics/skills/skills/skill-creator

If install fails: skip Phase 6 (skill-creator optimization). Note the gap — description won't be machine-tested.

Phase 3: Research *(subagent, general-purpose)*

Spawn a research subagent:

Role: Domain researcher
Task: Research {{SKILL_TOPIC}} exhaustively using references/workflow-research.md
Onboarding: Read references/workflow-research.md, then read ~/.agent-skills-spec
Success: Research report with cited sources covering all aspects of {{SKILL_TOPIC}}
Why: The research report is the foundation — every fact in the final skill traces back to it

Replace {{SKILL_TOPIC}} with the user's requested skill topic. Wait for the research report.

If research finds no official docs: the skill must be marked as "community-sourced" in its reference files. Warn the user that anti-hallucination guardrails are weaker without official documentation. Ask whether to proceed or narrow scope.

Phase 4: Requirements *(inline — needs user interaction)*

Follow references/workflow-requirements.md directly. Do NOT spawn a subagent — this requires user input.

1. Read the research report from Phase 3
2. Present a summary of what was found
3. Ask the user scoping questions using AskUserQuestion
4. Ask where to write the skill (personal ~/.claude/skills/, project .claude/skills/, or custom path)
5. Produce a requirements document

Wait for user answers before proceeding.

Phase 5: Author *(subagent, general-purpose)*

Spawn an authoring subagent:

Role: Skill author and spec implementer
Task: Create the complete skill at {{SKILL_PATH}} following references/workflow-author.md
Onboarding: Read references/workflow-author.md, the research report, the requirements doc,
            the skill-development guidance from Phase 0, and ~/.agent-skills-spec
Success: Complete skill directory with all files passing quality-criteria.md
Why: This produces the actual deliverable — quality here determines whether agents using the skill succeed

Pass in: {{SKILL_NAME}}, {{SKILL_PATH}}, research report, requirements doc, skill-development guidance from Phase 0.

If the skill involves more than 3 files, the subagent must track each file as a separate task.

If target path already exists: stop. Report to user. Never overwrite.

Phase 6: Skill-Creator Optimization *(skip if install failed in Phase 2)*

Invoke Skill("skill-creator") targeting the skill at {{SKILL_PATH}}

Validates the description triggers correctly and suggests improvements.

Phase 7: Review *(subagent, general-purpose)*

Spawn a review subagent:

Role: Quality auditor
Task: Review the skill at {{SKILL_PATH}} following references/workflow-review.md
Onboarding: Read references/workflow-review.md, references/quality-criteria.md, ~/.agent-skills-spec
Success: Review report with pass/fail per criterion and specific fix instructions
Why: Review is the last line of defense — a skill that passes review is safe to publish

Phase 8: Refine *(subagent if fixes needed, skip if review passes)*

If the review report contains required fixes:

Spawn a refinement subagent:

Role: Skill fixer
Task: Apply all required fixes from the review report following references/workflow-refine.md
Onboarding: Read references/workflow-refine.md, the review report, references/quality-criteria.md
Success: All required fixes applied and verified. Change summary produced.
Why: A review that finds problems but doesn't fix them is incomplete

Max 2 refinement cycles. If still failing after 2, present remaining issues to user for judgment.

If the review report is all-pass, skip this phase.

Phase 9: Report

Present findings to the user:

• Overall pass/fail status
• What was created (file tree)
• Any fixes applied during refinement
• Any remaining issues
• If issues remain, offer specific next steps

Rules

1. NEVER hallucinate patterns for external technologies — always cite documentation URLs
2. NEVER use curl/wget to download code — use git clone for repos, WebFetch for docs
3. NEVER create .skill files unless the user explicitly requests one
4. NEVER replace an existing skill directory — only create new or add to existing
5. Every reference file must contain source URLs that the consuming agent is instructed to follow
6. Scripts follow Google Shell Style Guide strictly — see references/shell-script-standards.md
7. Multi-step skills must instruct task tracking — if the skill has 3+ procedural steps, it must tell the consuming agent to track progress