create-skill

Files (1)

• SKILL.md

SKILL.md (first lines)

# Create Skill

Interactive wizard for creating a well-structured Claude Code skill in this repo.

## Behavior

If `$ARGUMENTS` is non-empty, use it as the proposed skill name (convert to kebab-case).
Otherwise, derive the name from the user's answers.

Run the interview in **three rounds**. Do not generate any files until round 3 is complete.

---

## Round 1 — Purpose and scope

Use `AskUserQuestion` with **these exact questions**:

**Question 1:** "What should this skill do? Describe the goal in one or two sentences."
- Header: "Purpose"
- Options: offer none (free text via Other)

Wait. Do not proceed until answered.

**Question 2:** "What type of skill is this?"
- Header: "Type"
- Options:
  - **Workflow** — Multi-step process with sequential steps, approval gates, or side effects (commit, deploy, publish)
  - **Knowledge** — Best practices, style guide, or conventions Claude should apply automatically
  - **Automation** — Runs a script or tool chain with minimal interaction. Triggered by user only.
  - **Hybrid** — Combines knowledge context with an explicit workflow

**Question 3 (multi-select):** "Which tools will this skill need? Select all that apply."
- Header: "Tools"
- Options:
  - Bash — run shell commands
  - Read / Write / Edit — work with files
  - Grep / Glob — search codebase
  - AskUserQuestion — ask the user questions during execution

---

## Round 2 — Invocation and triggers

After round 1, use `AskUserQuestion` with:

**Question 4:** "What phrases would a user say to trigger this skill? Give 2–4 examples of natural requests."
- Header: "Triggers"
- Options: offer none (free text)

**Question 5:** "Should Claude auto-invoke this skill, or should only the user invoke it manually?"
- Header: "Invocation"
- Options:
  - **Auto + manual** — Claude loads it when the request matches the description (default)
  - **Manual only** — Only runs when user explicitly types `/ccosming:skill-name`. Use for side-effect operations.
  - **Claude only, hidden** — Claude applies it automatically but it's hidden from the `/` menu. Use for background knowledge.

**Question 6:** "Does this skill take arguments? (e.g., a filename, issue number, or keyword)"
- Header: "Arguments"
- Options:
  - No arguments
  - Yes — one argument (describe it)
  - Yes — multiple named arguments (describe them)

---

## Round 3 — Workflow design

After round 2, ask:

**Question 7 (multi-select):** "Which of these does the skill need to handle?"
- Header: "Edge cases"
- Options:
  - Abort conditions (when to stop and tell the user)
  - Error handling (common failure modes with solutions)
  - Approval gate (propose → wait for user confirmation → execute)
  - Secrets / safety check before executing

**Question 8:** "Walk me through the main steps of this skill in order. List them briefly (one per line or comma-separated)."
- Header: "Steps"
- Options: offer none (free text)

---

## Generation — after all 8 questions are answered

### 1. Determine skill name

- If `$ARGUMENTS` is non-empty → use as `SKILL_NAME` (convert spaces to dashes, lowercase).
- Otherwise → derive a kebab-case name from the purpose answer (max 3 words).

Confirm: "I'll name this skill `{SKILL_NAME}`. The command will be `/ccosming:{SKILL_NAME}`. Continue?"
Wait for approval. Adjust name if user requests.

### 2. Build frontmatter

Construct YAML frontmatter:

```yaml
---
name: {SKILL_NAME}
description: >
  {What it does — one sentence}. {When to use it — include 2-3 trigger phrases from Q4}.
  {Key capability or output}.
license: MIT
allowed-tools: {comma-separated tools from Q3}
argument-hint: "{hint from Q6, if applicable}"
arguments: [{arg1}, {arg2}]          # only if Q6 said multiple args
disable-model-invocation: true        # only if Q5 = "Manual only"
user-invocable: false                 # only if Q5 = "Claude only, hidden"
---
```

Rules:
- `description` ≤ 1024 characters. Must include trigger phrases from Q4.
- `allowed-tools` only if tools were selected in Q3.
- Omit `argument-hint` and `arguments` if Q6 = "No arguments".
- Omit `disable-model-invocation` and `user-invocable` unless explicitly required.

### 3. Build body

From ccosming/skills, MIT.

More from ccosming/skills