/bench-new-skill — Create New Skills

Scaffold new skills following established conventions, the official Claude Code skills spec, and Anthropic's Complete Guide to Building Skills.

When to Use

• User wants to create a new skill (anywhere — global, project-local, or bench-skills repo)
• User says "create a new skill", "add a skill", "new slash command"
• User wants to create a plugin (see references/plugin-guide.md)

Principles

1. Start with use cases, not code. Define 2-3 concrete use cases before writing anything. See references/anthropic-guide-insights.md for the use case template.
2. Description = [What it does] + [When to use it] + [Key capabilities]. Must include BOTH what and when. Under 1024 chars. No XML tags. Include specific trigger phrases users would say.
3. Include only what Claude doesn't already know. Don't teach Claude how to write code — teach it your specific process, conventions, and standards.
4. Be specific and actionable. Vague instructions like "validate the data" get ignored. Explicit instructions like "Run scripts/validate.py --input {file}" get followed.
5. Self-contained skills. Each skill directory must work independently when installed.
6. Progressive disclosure. Keep SKILL.md focused on core instructions (under 300 lines / 5,000 words). Move detailed docs to references/ and link to them.
7. Right invocation model. Decide upfront: should the user invoke this, Claude invoke it, or both? Use disable-model-invocation and user-invocable accordingly.