/skill-creator

name: skill-creator
description: "Creates new skills from templates via guided workflow. Triggers: new skill, create skill, skill scaffold, skill template."
effort: high
disable-model-invocation: true
argument-hint: "[skill name or description]"
allowed-tools: Read, Write, Edit, Bash, Grep, Glob

Skill Creator

$ARGUMENTS

Create a new skill following the Agent Skills standard.

Workflow

1. **Capture intent** -- ask: what should the skill do? Who invokes it? 2. **Classify** -- task, hybrid, or knowledge? (see Classification Guide below) 3. **Interview** -- gather: framework, tools, scope, output format, constraints 4. **Write SKILL.md** -- frontmatter + body, under 500 lines 5. **Create supporting files** -- reference/, templates/, scripts/ as needed 6. **Test and iterate** -- invoke, observe, refine

Frontmatter Reference

This table describes Claude Code skill fields. ai-toolkit requires `name` and `description` for portable catalog entries even though Claude Code can infer them.

Only set `model` or `effort` when the skill deliberately needs an approved override. Omission inherits the session choice in Claude Code. Check supported effort levels for the selected model; these fields do not configure Codex or Copilot reasoning. Refer to `kb/reference/model-compatibility.md` for the current client/API boundaries and verify actual account availability before pinning IDs.

| Field | Type | Required | Description | |-------|------|----------|-------------| | `name` | string | yes | Lowercase, hyphens only, max 64 chars | | `description` | string | yes | Third person, max 1024 chars, include key terms | | `effort` | low/medium/high/xhigh/max | no | Override effort for the invoking turn; supported levels depend on the model | | `disable-model-invocation` | bool | no | `true` = only user can trigger (task skills) | | `user-invocable` | bool | no | `false` = knowledge skill, Claude auto-loads | | `allowed-tools` | string or YAML list | no | Pre-approve tools for the invoking turn; does not restrict the available tools | | `disallowed-tools` | string or YAML list | no | Remove tools for the invoking turn | | `model` | string | no | Override the model for the invoking turn, or use `inherit` | | `context` | string | no | `fork` to run in isolated subagent | | `agent` | string | no | Agent type to use when `context: fork` | | `argument-hint` | string | no | Shown in autocomplete, e.g. `"[target]"` | | `arguments` | string or YAML list | no | Named positional arguments, substituted as `$name` in the body | | `when_to_use` | string | no | Additional trigger context appended to the description | | `background` | bool | no | With `context: fork`, `false` waits for the subagent result; default `true` | | `hooks` | object | no | Hooks registered at invocation and retained for the session; `once` limits a handler to one run | | `paths` | string or YAML list | no | File globs limiting automatic activation | | `shell` | bash/powershell | no | Shell for dynamic context commands | | `metadata` | object | no | Custom metadata for external tooling | | `license` | string | no | Skill license | | `compatibility` | string | no | Runtime requirements, up to 500 characters |

`disallowedTools`, `maxTurns`, `memory`, and `skills` belong to agent frontmatter, not skill frontmatter. Put them on the agent referenced by a forked skill. Direct claude.ai skill uploads and the Skills API accept only `name`, `description`, `license`, `compatibility`, `metadata`, and `allowed-tools`; Claude Code plugin skills support the full table. See the [official skill reference](https://code.claude.com/docs/en/skills#frontmatter-reference) (reviewed 2026-09-23).

Classification Guide

| Type | When to use | Key fields | |------|-------------|------------| | **Task** | User-triggered actions with side effects (build, deploy, commit) | `disable-model-invocation: true`, `allowed-tools` | | **Hybrid** | Both user and Claude invoke (review, analyze) | defaults | | **Knowledge** | Domain patterns Claude auto-loads (clean-code, api-patterns) | `user-invocable: false` |

Writing Guidelines

• **Description**: third person ("Generates...", "Provides..."), **min 50 chars**, include searchable key terms **and a trigger hint** (e.g., "Use when..."). Short descriptions cause over-triggering — adjacent skills fight for the same query.
• **Name**: lowercase, hyphens, max 64 chars -- match the directory name
• **Length**: SKILL.md under 500 lines; use `reference/` for overflow
• **Be concise**: Claude is smart -- give structure, not lectures
• **Degrees of freedom**: be specific on format/constraints, flexible on implementation
• **File references**: max 1 level deep (e.g., `reference/details.md`)
• **Use `$ARGUMENTS`**: place it early so user input is visible
• **Tables over prose**: for options, patterns, mappings

Mandatory Sections (meta-architect audit criteria)

Every SKILL.md must contain all five, or it will lower the toolkit quality score:

1. **Description ≥50 chars** with trigger hint in frontmatter 2. **At least one concrete example** — fenced code block or `## Example` section 3. **Hard rules** using uppercase markers: `MUST`, `NEVER`, `CRITICAL`, `MANDATORY`, or `DO NOT` 4. **"When NOT to Use" section** — list 2-3 adjacent skills and the boundary between them 5. **Under 500 lines** — split into `reference/` if needed

These map directly to the mutation strategies in `meta-architect.md`: `add_example`, `add_constraint`, `add_edge_case`, `restructure`.

Rules vs Gotchas — different semantic buckets

Split the "hard rules" criterion into two sections when both apply:

• **`## Rules`** — prescriptive process constraints (always-true MUST / NEVER): *"MUST form a testable hypothesis before changing code"*, *"NEVER force-push main"*. Required in every skill.
• **`## Gotchas`** — environment-specific traps the agent would miss without being told, taken from [Anthropic's best practices](https://agentskills.io/skill-crea