| name | create-skill |
| description | Create a new Claude Code skill from a description. Use when the user says "create skill", "new skill", or "add a slash command". |
| argument-hint | <skill-name> |
| user-invocable | true |
| allowed-tools | Read, Write, Glob |
| model | sonnet |
/create-skill — Generate a new skill
Create a new skill (slash command) following the framework's conventions.
Steps
-
If $ARGUMENTS is empty, ask the user for:
- Skill name (lowercase, hyphenated)
- What the skill should do (one sentence)
- When it should trigger (what the user would say)
-
Read an existing skill for reference patterns:
.claude/skills/review/SKILL.md
-
Generate the skill file following this structure:
---
name: <skill-name>
description: <what it does>. Use when the user says "<trigger phrases>".
argument-hint: "<hint>"
user-invocable: true
allowed-tools: <minimal set needed>
model: <haiku for simple, sonnet for complex>
---
Body structure:
# /<skill-name> — <one-line description>
<What this skill does, in one sentence.>
## Steps
1. <first action>
2. <second action>
3. <report results in specific format>
## Arguments
- No arguments: <default behavior>
- `$ARGUMENTS`: <behavior with arguments>
## Notes
- <important constraint 1>
- <important constraint 2>
-
Write the skill to .claude/skills/<skill-name>/SKILL.md
-
Report what was created:
Created: .claude/skills/<skill-name>/SKILL.md
Trigger: /<skill-name>
Description: <what it does>
Guidelines
- Context efficiency: Keep skills under 100 lines. The context window is a shared resource.
- Model selection: Use
haiku for simple tasks (build, test, file checks). Use sonnet for tasks requiring analysis (review, planning).
- Tool minimization: Only include tools the skill actually needs in
allowed-tools.
- No side effects: Skills should report findings — not silently make changes unless that's their explicit purpose.
- Follow conventions: Match the tone and format of existing skills in the project.
