What Counts as a Skill: SKILL.md, Frontmatter, Scripts

A Funplay skill is a directory under skills/ containing a required SKILL.md file and an optional scripts/ folder. The SKILL.md frontmatter declares the skill's name, description, dependencies, inputs, and outputs—everything a coding agent needs to discover and execute the workflow without runtime tooling.

What Funplay Skills Are and Why They Matter

The funplay-skill package (version 0.1.0, MIT-licensed) is a collection of twelve reusable, deterministic AI workflows packaged as plain markdown and optional scripts. It lives at https://github.com/FunplayAI/funplay-skill, is maintained by FunplayAI, and requires Node.js 18 or later with [email protected] as its package manager. The repository is inspired by obra/superpowers but narrows its focus to game-development art, audio, engine workflows, and asset processing.

Skills are not MCP servers. An MCP server exposes callable tools at runtime over HTTP—live endpoints you hit to get things done. A skill is a static document that teaches a coding agent what to do. The agent reads the markdown, understands the workflow, and then uses its existing file and shell tools to execute the steps. The two approaches compose cleanly: you can run Funplay's Unity MCP for hands-on editor control and Funplay Skills for high-level playbooks that orchestrate multi-step processes.

The package manifest declares a contextFileName of GEMINI.md, which tells compatible coding agents where to look for skill context. Agents discover available skills through the plugin manifest registered for their client. The skill inventory spans nine domains: five are engine-agnostic (covering art, audio, design, and onboarding), three are engine-specific (unity-prefab-workflow for Unity, godot-scene-assembly for Godot, and cocos-component-workflow for Cocos Creator), and one (using-funplay-skills) is a meta skill that guides users through discovering, choosing, and applying the other eleven.

Anatomy of a Skill: SKILL.md, Frontmatter, and Scripts

Every skill follows the same directory convention:

skills/
  my-skill-name/
    SKILL.md        ← required
    scripts/        ← optional
      setup.sh
      validate.sh

The SKILL.md file is the contract. Its frontmatter block is parsed by the agent to understand what the skill needs and what it produces. Here is the top-level package manifest for context:

{
  "name": "funplay-skill",
  "description": "Game-development skills library for coding agents.",
  "version": "0.1.0",
  "contextFileName": "GEMINI.md"
}

Each of the twelve skills contains its own SKILL.md with frontmatter declaring five fields:

Consider the unity-prefab-workflow skill. Its purpose is to safely plan and review Unity prefab, scene, and serialized asset edits. It targets Unity specifically and carries no external dependencies beyond the engine itself. The godot-scene-assembly skill organizes Godot scenes, nodes, and resources cleanly. The cocos-component-workflow skill plans and reviews Cocos Creator prefab, component, and asset reference changes. All three follow the same SKILL.md + optional scripts/ pattern.

The scripts/ directory is where determinism lives. When a skill includes scripts, those scripts encapsulate operations that must produce identical results given identical inputs—no ambiguity, no LLM hallucination in the execution layer. The agent reads SKILL.md to understand the workflow, then invokes scripts from scripts/ using its shell tooling. The markdown body below the frontmatter contains the prose instructions: context, edge cases, validation steps, and examples.

Engine-agnostic skills follow the same structure but omit engine-specific dependencies. A texture-processing skill under the art domain, for example, declares its input formats and output conventions in frontmatter, provides validation scripts if needed, and documents the expected pipeline in the markdown body. The agent does not need to know whether the final destination is Unity, Godot, or Cocos Creator—the skill handles the abstraction.

The meta skill, using-funplay-skills, serves as the onboarding entry point. It teaches the agent how to discover available skills, evaluate which one fits the current task, and apply it correctly. Think of it as the skill that teaches agents how to use skills—a recursive pattern that keeps the system self-documenting.

Gotchas and Limitations

Skills are static. They cannot respond to runtime state the way an MCP server can. If your workflow requires live editor queries—reading the current scene hierarchy in Unity, polling a running game process, or inspecting a runtime asset database—you need an MCP server for that layer. Skills tell the agent the playbook; MCP servers give the agent hands.

The funplay-skill package is at version 0.1.0 and under active development. APIs, frontmatter schemas, and directory conventions may change. Pin your dependency versions if you integrate this into a production pipeline.

Node.js 18 is the minimum requirement, and the repo recommends running npx pnpm if you do not have pnpm installed globally. If your environment runs an older Node version, the package will not install correctly.

Skills do not ship with a runtime server. There is no HTTP endpoint, no 127.0.0.1:8765 service to start, no socket to connect to. The entire value proposition lives in the markdown and scripts that coding agents read and execute using their own tool chains. If you are expecting a daemon process or a background service, you are looking for the MCP server repositories, not the skills repository.

Finally, engine-specific skills are opinionated. The unity-prefab-workflow skill assumes Unity's serialization model and YAML-based prefab format. The godot-scene-assembly skill assumes Godot's .tscn structure. If your project uses a custom pipeline that diverges significantly from the engine's defaults, you may need to fork and modify the skill rather than use it as-is.

Closing Thoughts

Funplay Skills gives coding agents a structured playbook format for game-development workflows—plain markdown with optional scripts, discoverable through plugin manifests, and composable with existing MCP tooling. Browse the full skill inventory and contribute your own at the funplay-skill repository. For deeper integration guides and workflow examples, visit https://gamebooom.ai/en/blog/. If you need live editor tooling alongside these playbooks, explore the sibling repositories under the FunplayAI organization, such as the FunplayAI Unity MCP server for hands-on Unity control.