Skill: Skill Creator
Purpose: Create a new skill using the standard system template.
When: Improve Skill proposes a new skill, or user wants to automate a recurring action.
Output: Ready ~/.claude/skills/[name]/SKILL.md + row in ~/.claude/skills/index.md.
PROCESS
Step 1 — Ask 3 questions
Создаём новый скилл. Три вопроса:
1. НАЗВАНИЕ и назначение?
Название — kebab-case (например: code-review, deploy, test-runner)
Назначение — одно предложение: что делает и когда использовать
2. ТРИГГЕР?
При каком условии этот скилл должен запускаться?
3. ЕСТЬ ЛИ ПОХОЖИЙ СКИЛЛ в ~/.claude/skills/index.md?
Проверь индекс — если есть похожий, лучше расширить его
Step 2 — Check for duplicates
Read ~/.claude/skills/index.md.
If similar skill exists — offer to extend it instead:
Скилл [existing] уже делает [similar thing].
Предлагаю добавить [new behavior] в него, а не создавать отдельный.
Согласен?
Step 3 — Check cross-cutting principles
Read ~/.claude/skills/cross-cutting-principles.md.
Verify the new skill will satisfy every principle before generating.
Flag any principle that the skill design might violate — resolve before proceeding.
Step 4 — Generate SKILL.md
Create file ~/.claude/skills/[name]/SKILL.md using template:
# Skill: [Name]
**Purpose:** [one sentence]
**When:** [trigger]
**Output:** [concrete artifact or action]
---
## PROCESS
### Step 1 — [First action]
[What agent does]
### Step 2 — [Next step]
[What agent does]
---
## OUTPUT
[Result template or description of expected output]
---
## CHECKLIST
- [ ] [quality criterion]
- [ ] [quality criterion]
---
## PRE-FLIGHT
Before delivering output, re-read ## RULES and ## CHECKLIST.
Verify output complies with every item. Fix before delivery.
Step 5 — Update ~/.claude/skills/index.md
Add row in the correct group table:
| [name] | [purpose] | [tags] | 0 | ✅ Ready |
Tags: meta loop start tools code test deploy review security data
Step 6 — Show result
Created: ~/.claude/skills/[name]/SKILL.md
Added to: ~/.claude/skills/index.md
Invoke: "@~/.claude/skills/[name]/SKILL.md" or "use skill [name]"
Optional: create ~/.claude/commands/[name].md?
RULES
- One skill = one responsibility. Don't combine unrelated actions.
- Name in lowercase, kebab-case.
- Process must be step-by-step — agent must know what to do at each step.
- Output template required if skill creates a file.
- Checklist before final — minimum 2 criteria.