# Skill Creator

> Meta-skill for authoring Markdown Skill files with YAML frontmatter and progressive disclosure via create_skill. Use when: adding a new skill, generating SKILL.md with references or templates, or checking description rules.

- Skill: `xuiltul/skill-creator` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add xuiltul/skill-creator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xuiltul/skill-creator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Docs & Writing
- Author: xuiltul (https://skillmd.com/u/xuiltul)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/xuiltul/skill-creator

---


# skill-creator

## Skill file structure

A Skill file consists of YAML frontmatter and Markdown body.
Required frontmatter fields are `name` and `description`.
Optional metadata may include tool constraints, trust/provenance fields, category, prompt policy, and routing hints. Use optional fields only when they help selection, safety, or maintenance.

```yaml
---
name: skill-name
description: >-
  Concise third-person summary of what the skill does.
  Use when: comma-separated scenarios where this skill applies.
allowed_tools:
  - read_memory_file
trust_level: trusted
source:
  type: anima
  origin: manual
category: communication
use_when:
  - drafting partner emails
trigger_phrases:
  - draft a partner email
negative_phrases:
  - personal diary
domains:
  - gmail
routing_examples:
  - Prepare a reply draft for the bank thread
---
```

`description` is the primary field for discovery and selection: the model uses it to decide relevance.
The body is read when you open the path from the system prompt skill catalog with `read_memory_file` (e.g. `skills/foo/SKILL.md`, `common_skills/bar/SKILL.md`).

Supported `create_skill` optional arguments include `references`, `templates`, `allowed_tools`, `trust_level`, `source_type`, `source_origin`, `category`, `promotion_status`, `skill_policy`, `use_when`, `trigger_phrases`, `negative_phrases`, `domains`, and `routing_examples`.

**Authoring format**: follow **`Use when:`** as described in `references/description_guide.md` (Agent Skills standard).
After editing, validate with **`python scripts/lint_skill.py path/to/SKILL.md`**.

## Writing `description`

Do not use legacy **`「」` keyword lists**. Use a short third-person capability line plus **`Use when:`** with concrete verbs and nouns.

See **`references/description_guide.md`** for rules (250 characters, no XML tags, examples, checklist).

### Domain-specific and concrete

Generic wording causes false positives. Prefer tool names, operations, and targets specific to the skill.

## Progressive disclosure

Skill information is disclosed in three levels.

| Level | Content | When shown |
|-------|---------|------------|
| Level 1 | `name` + `description` | Skill catalog / tool descriptions (budgeted) |
| Level 2 | body | Loaded with `read_memory_file(path="skills/.../SKILL.md")` or `common_skills/.../SKILL.md` |
| Level 3 | External files | Loaded per body instructions (`references/`, `templates/`) |

Keep Level 1 concise; put procedures in Level 2; offload long material to Level 3.

## Creation procedure

### Step 1: Clarify

- What to automate or document
- Personal vs common Skill (procedures use `procedures/*.md` separately)
- **Use when:** scenarios (when to choose this skill)

### Step 2: Design

- **name**: kebab-case (e.g. `my-skill`); use `*-tool` naming for external tool guides when applicable
- **description**: third-person summary + **`Use when:`** (see `references/description_guide.md`)
- **body**: section structure; optional `{{now_local}}` and other builtins
- **references** / **templates**: optional
- **allowed_tools**: optional soft constraint
- **trust/source/category/policy/routing**: optional trust level, provenance, category, prompt policy, and routing metadata (`use_when`, `trigger_phrases`, `negative_phrases`, `domains`, `routing_examples`)

### Step 3: Create

```
create_skill(skill_name="{name}", description="{description}", body="{body}")
```

Common skills:

```
create_skill(skill_name="{name}", description="{description}", body="{body}", location="common")
```

You can also pass `references`, `templates`, `allowed_tools`, `trust_level`, `source_type`, `source_origin`, `category`, `promotion_status`, `skill_policy`, `use_when`, `trigger_phrases`, `negative_phrases`, `domains`, and `routing_examples` when those fields are useful.

Prefer `create_skill` for new skills; flat `skills/foo.md` alone may not match `skills/foo/SKILL.md` for `read_memory_file`.

### Step 4: Verify

- Re-read with `read_memory_file(path="skills/{name}/SKILL.md")` (or the `common_skills/...` path from the catalog)
- Run **`python scripts/lint_skill.py`** on the file (recommended)

## Checklist

- [ ] YAML frontmatter delimited by `---`
- [ ] `name` and `description` present
- [ ] **`Use when:`** present; no **`「」`** keyword enumeration
- [ ] Domain-specific, concrete wording (avoid vague “manage” / “check” alone)
- [ ] Body has actionable steps
- [ ] Avoid relying only on `## Overview` for description; prefer frontmatter
- [ ] Optional metadata (`trust_level`, `source`, `category`, `skill_policy`, `use_when`, `trigger_phrases`, `negative_phrases`, `domains`, `routing_examples`) matches the actual skill
- [ ] Created via `create_skill` with `{name}/SKILL.md` layout where applicable

## Template

Use `templates/skill_template.md` bundled with this skill, or:

```markdown
---
name: {{skill_name}}
description: >-
  {{Line 1: concise capability summary}}
  Use when: {{comma-separated usage scenarios}}
---

# {{skill_name}}

## Procedure

1. ...
2. ...

## Do not place symlinks in skill trees

Never "re-publish" a host-side skill (e.g. one under `~/.claude/skills`) into `common_skills/` or `skills/` via a symlink. `read_memory_file` checks the resolved real path, so a symlink that points outside the tree fails with "Path traversal detected", and the catalog lists the symlink as a native skill that shadows the readable `external/<engine>/<name>/SKILL.md` entry. Host-side skill roots are injected automatically via `skills.external_roots`; if a real copy is needed, copy the file and decide which side is canonical.

## Notes

- ...
```

## Notes

- Skills are Markdown playbooks, not Python tools
- Required frontmatter: `name`, `description`
- Optional: `allowed_tools`, trust/provenance fields, `category`, `skill_policy`, and routing metadata. Keep metadata minimal when the description is sufficient
- Keep body around 150 lines when practical; use `references/` for long material