skill-creator（技能创建器）

Skill Creator

Build skills that behave like “mini senior engineers”: trigger correctly, run deterministically, and leave resumable artifacts on disk.

What good looks like (Ship Faster standard)

• SKILL.md: short and triggerable (routing + safety + I/O + output contract). Avoid long tutorials here.
• references/: deep docs loaded only when needed (progressive disclosure). Link directly from SKILL.md.
• scripts/: deterministic helpers for repetitive/fragile work (prefer this over re-writing code in chat).
• assets/: templates or files to copy (not meant to be loaded into context).

Deep guidance:

• Progressive disclosure + authoring principles: references/skill-design-principles.md
• Workflow structure patterns: workflows.md
• Output templates/patterns: output-patterns.md

Creation flow (recommended)

1. Collect trigger examples (real user phrases that should activate the skill).
2. Define the I/O contract (paths only) + required artifacts:

- What files does the skill read? - What files must it write (plans, evidence, summaries)?

1. Implement reusable resources first:

- add scripts/ for deterministic steps - add references/ for heavy documentation - add assets/ for templates / boilerplate

1. Write SKILL.md last:

- frontmatter description should include trigger phrases - body should be short and link out to the resources above

1. Validate + package (optional) using the scripts below.

Scripts

All scripts live in scripts/ in this skill folder.

Initialize a new skill folder

python scripts/init_skill.py <skill-name> --path <output-directory>

Validate a skill (frontmatter + basic conventions)

python scripts/quick_validate.py /path/to/<skill-name>

Sync catalogs/assets (manifest, skills-map, docs catalog)

python scripts/sync_catalog.py

Lint skills (compat + quality gates)

python scripts/skill_lint.py --check-generated

Package a skill into a distributable .skill zip (optional)

python scripts/package_skill.py /path/to/<skill-name> [output-directory]

Notes for Ship Faster contributors

• Prefer prefix naming: workflow-, tool-, review-, mcp-, skill-, publish-.
• Keep side effects gated: write a plan first, ask for explicit confirmation, then execute, then verify.
• If you add/remove a skill, update the catalog: skills/manifest.json.

Claude Code: Slash Commands merged into Skills

Claude Code now treats custom slash commands and skills as the same concept:

• .claude/commands/<name>.md and .claude/skills/<name>/SKILL.md both create /<name>
• Existing .claude/commands/ files keep working (no migration required)
• Prefer creating skills going forward so you can bundle references/, scripts/, and use subagents

Useful frontmatter knobs (Claude Code):

• disable-model-invocation: true - only user can invoke (good for side-effect workflows)
• user-invocable: false - hide from users; Claude-only background knowledge
• context: fork + agent: Explore|Plan|... - run skill in a subagent (good for research)
• allowed-tools - restrict tool surface when this skill is active

Archived full guide

The previous (long-form) version of this skill is preserved at:

• references/legacy-skill-creator.md