--- name: skill-creator description: > Create high-quality AI Skills (knowledge, tool, or hybrid) using the local standards and templates. Trigger: Use when the user asks to create a new skill, package a repeated pattern (≥ 3 uses), standardize a workflow, or generate a reusable rule set. license: Apache-2.0 metadata: type: hybrid author: devcontext version: "2.4.0" scope: [root] auto_invoke: false # Meta-skill: Manual invocation preferred with @skill-creator allowed-tools: [Read, Write, Bash] --- # Skill Creator 🧱 ## TL;DR (REQUIRED) * **Before creating any skill**, you MUST read: `reference\/skill-designer-core.md` and `assets\/SKILL-TEMPLATE.md`. * You MUST output **Creator compliance** as the **first visible output**. * Keep scope tight: **one skill = one job**. If it mixes responsibilities → **split**. --- ## 🚨 Mandatory reading (Non-negotiable) You MUST read these files in this order before generating anything: 1. `reference\/skill-designer-core.md` 2. `assets\/SKILL-TEMPLATE.md` If you haven’t read them, stop and say: ``` Skill creation blocked: designer core and template not loaded. ``` --- ## 🧾 Creator compliance (MANDATORY) You MUST output this block as the **first visible output** (before any narration, planning, tasks, or file edits): ``` Creator compliance: - Reusability (≥ 3 uses): YES\/NO + 1-line justification - Skill type: knowledge | tool | hybrid - Single responsibility: - Scope boundaries: - In scope: - Out of scope: - Cross-references (do not duplicate): - Split decision: NONE | SPLIT INTO , ``` Rules: * If Reusability = NO → do not create a skill. Provide a snippet\/checklist instead. * If Split decision ≠ NONE → create multiple skills instead of one. --- ## When to use Use this skill to transform a request into a structured skill folder when: * A pattern repeats **≥ 3 times**. * The project needs explicit guardrails for AI behavior. * The request is deterministic enough to package (knowledge and\/or automation). Do NOT use when: * It’s a one-off task (make a snippet instead). * The documentation already exists (reference it). * The scope is too broad (split first). --- ## Outputs (what you are allowed to change) ### ✅ Allowed * Create a new skill folder under the local skills directory, e.g.: * `.agent\/skills\/\/...` (preferred in consumer repos) * `skills\/\/...` (when the repo uses `skills\/`) ### ❌ Not allowed * Creating repository-level planning artifacts like `task.md`, `implementation_plan.md`, or similar. * Modifying product\/app source code while “creating a skill”. * Modifying other skills unless explicitly requested. * Adding external URLs inside `reference\/` files. --- ## Critical patterns ### ALWAYS * **Reusability filter**: confirm it will be used ≥ 3 times. If not, reject. * **Single responsibility**: one skill = one job. Split monoliths. * **Progressive disclosure**: keep `SKILL.md` < 500 lines; move deep content to `reference\/`. * **Explicit trigger**: description MUST include an actionable Trigger. * **Minimum permissions**: request only the tools needed for the skill type. * **Cross-references**: point to existing skills instead of duplicating their scope. **Hard split rule (common failure mode):** * If a skill mixes **framework mechanics** (e.g. Next.js routing\/runtime) and **project policy\/contract** (e.g. API envelopes, auth rules, CORS policy), you MUST split into separate skills. ### NEVER * Define component structure (belongs to `react-19`). * Define file\/folder placement decisions (belongs to `structuring-projects`). * Define styling patterns (belongs to `tailwind-4`). * Duplicate content between `SKILL.md` and `reference\/`. * Use external URLs in `reference\/` (copy locally). --- ## Decision tree (skill type) ``` Does it only teach rules\/patterns? → Knowledge skill (allowed-tools: Read) Does it automate a deterministic task? → Tool skill (allowed-tools: Read, Write, Bash) Does it teach rules AND automate? → Hybrid skill (allowed-tools: Read, Write, Bash) ``` --- ## Workflow (Factory Process) ### Phase 1 — Validate * Print Creator compliance. * Clarify the job-to-be-done. * Confirm ≥ 3 uses. * Enforce single responsibility (split if needed). ### Phase 2 — Scaffold Create: * `\/SKILL.md` * `\/reference\/` * `\/assets\/` * `\/scripts\/` (tool\/hybrid only) ### Phase 3 — Draft * Populate `SKILL.md` from `assets\/SKILL-TEMPLATE.md`. * Keep the body as an orchestrator: TL;DR → When to Use → Outputs → Scope → Critical Patterns → Decision Tree → Actions (if tool\/hybrid) → Cross-references → Resources. * Add cross-references instead of duplicating other skills. ### Phase 4 — Quality Gate (MANDATORY) Run the checklist from `reference\/skill-designer-core.md`. Additionally, enforce these objective checks (prefer via Bash): * **No external URLs in `reference\/`** (reject if found): ```bash grep -RInE "https?:\/\/" reference\/ && echo "FAIL: external URLs found in reference\/" && exit 1 || true ``` * **`SKILL.md` length** (< 500 lines): ```bash wc -l SKILL.md ``` If any item fails, fix it before completion. ### Phase 5 — Activation hint When finished, recommend using `skill-integrator` to activate the new skill in `AGENTS.md`. --- ## Resources * **Design Standards**: `reference\/skill-designer-core.md` * **Master Template**: `assets\/SKILL-TEMPLATE.md`