skill-creator · diff
git:20260321.8fae238 to git:20260321.9fa274a
45 added, 12 removed. Audit A to A.
---
name: skill-creator
description: Design a new Claude Code skill from a concrete use case and produce a repo-ready skill package. Use when the user asks to create a new skill, define a skill's responsibility, draft SKILL.md, choose frontmatter, design supporting files, or prepare eval criteria for a new skill. Also trigger on "スキルを作りたい", "スキルを作って", "スキルを追加して", "新しいスキル", "SKILL.md生成".
allowed-tools: Read, Grep, Glob, Bash, Write, Edit
---
+ ## Pattern declaration
+
+ Primary pattern: Pipeline
+ Secondary patterns: Inversion
+ Why: skill creation requires strict phase order (requirements → pattern selection → structure → review) with an upfront question-gathering gate.
+
## Purpose
Create a new skill only after the use case is concrete enough.
Your job is to:
1. identify the smallest useful responsibility
- 2. define when the skill should trigger
- 3. choose safe frontmatter defaults
- 4. draft a minimal but strong `SKILL.md`
- 5. propose supporting files only when they add clear value
- 6. define an initial eval set
+ 2. select the right design pattern
+ 3. define when the skill should trigger
+ 4. choose safe frontmatter defaults
+ 5. draft a minimal but strong `SKILL.md`
+ 6. propose supporting files only when they add clear value
+ 7. define an initial eval set
## Safety constraints
Write output files only to `.claude/skills/` unless the user explicitly specifies a different path.
Do not modify existing skills without user confirmation.
Note: this skill does not set `disable-model-invocation: true` because file creation is its primary purpose, not an incidental side effect. The settings.json deny list and user confirmation prompts provide sufficient guardrails.
## Core rule
Do not create a broad or fuzzy skill.
Prefer:
- one responsibility
- one primary workflow
- one clear success condition
If the requested scope is too broad, split it into multiple skills.
## Phase 0: Gate
Before drafting, confirm:
- target outcome
- primary user intent
- trigger phrases
- audience
- allowed tools needed
- whether the skill should be auto-invoked or manual-only
- source material available
- definition of done
If any item is unclear:
- ask the minimum focused questions, or
- inspect available files first if the answer can be found without asking
Do not draft the full skill until the gate passes.
+ ## Phase 0.5: Pattern selection
+
+ Choose the primary design pattern before drafting the skill.
+
+ Available patterns:
+
+ - Tool Wrapper — the skill mainly injects domain, library, or framework knowledge on demand
+ - Generator — the skill mainly produces structured output from a reusable template
+ - Reviewer — the skill mainly evaluates output against a checklist or rubric
+ - Inversion — the skill must gather missing information before acting
+ - Pipeline — the skill must enforce a strict multi-step workflow with checkpoints
+
+ Selection rules:
+
+ 1. identify the primary value source or failure mode the skill addresses
+ 2. map it to one primary pattern
+ 3. add secondary patterns only when one pattern is insufficient
+ 4. each secondary pattern must appear in a named phase with a measurable role
+
+ Do not continue until the pattern choice is explicit.
+
## Phase 1: Skill boundary design
Produce a short working summary with:
- skill name candidate
- single responsibility
- in-scope tasks
- out-of-scope tasks
- likely trigger language
- risk level
Reject designs that combine unrelated workflows.
## Phase 2: Frontmatter design
Choose frontmatter intentionally.
Defaults:
- include `name`
- include `description`
- omit extra fields unless clearly needed
Use:
- `disable-model-invocation: true` for side-effecting, sensitive, or timing-dependent workflows
- `allowed-tools` only when safe pre-approval is clearly useful
- `model` only when there is a concrete reason to override the default
For every chosen field, state why it is needed.
## Phase 3: Structure design
Draft the skill with this order:
- 1. purpose
- 2. gate
- 3. workflow phases
- 4. output contract
- 5. review step
- 6. failure handling
- 7. additional resources
+ 1. pattern declaration
+ 2. purpose
+ 3. gate
+ 4. workflow phases
+ 5. output contract
+ 6. review step
+ 7. failure handling
+ 8. additional resources
Keep the main file compact.
Move heavy references, examples, and templates into separate files.
Use `${CLAUDE_SKILL_ROOT}/assets/basic-skill-template.md` as the starting skeleton.
## Phase 4: Eval design
Create an initial eval plan with:
- 10 to 20 realistic test prompts
- 3 to 6 pass/fail criteria
- likely failure modes
- what should be checked by code vs by reviewer
Include at least:
- one trigger test
- one under-specified request
- one boundary case
- one anti-pattern case
Use `${CLAUDE_SKILL_ROOT}/assets/eval-rubric-template.md` as the rubric structure.
## Phase 5: Review
Check the draft against:
- `${CLAUDE_SKILL_ROOT}/references/review-default.md`
- `${CLAUDE_SKILL_ROOT}/references/design-principles.md`
Fix any issue before finalizing.
## Phase 6: Registration
After the skill files are written:
1. verify the directory name matches the `name` field in frontmatter
2. verify `description` is under 1024 characters and includes trigger keywords
3. verify SKILL.md is under 500 lines
4. verify all `${CLAUDE_SKILL_ROOT}/` references point to existing files
5. check the repository CLAUDE.md or AGENTS.md for additional registration steps
6. if the repository uses a skills index or registry, update it
## Output format
Return exactly these sections:
### Skill summary
+ ### Pattern selection
+
### Recommended directory structure
### SKILL.md draft
### Supporting files to add
### Eval plan
### Known risks
## Anti-patterns
Do not:
- create a "do everything" skill
- invent company-specific rules
- write a vague description
- skip the gate
+ - skip pattern selection
- omit eval planning
- hide trade-offs
+ - stack multiple patterns without justifying each one
## Failure handling
If the request is too broad:
- propose 2 to 4 smaller skills
- explain the split by responsibility
- recommend which one to build first