What This Project Does

scenario-labs/skills is a set of Agent Skills for AI coding agents. It packages the work of producing production grade creative content with AI into a set of reusable SKILL.md skill files that follow the open Agent Skills format. More than 70 agents can use it, including Claude Code, Cursor, Codex, and Copilot.

The creative content it can generate is broad: images, video, audio, textures, skyboxes, 3D assets, and custom trained models, aimed mainly at games, entertainment, and all kinds of creative work.

The skills come in two families. The first is Scenario skills, which call the cloud through the Scenario MCP server to generate content. The second is Expert Tools, where one lead skill teams up with several specialist skills to drive the DCC software and game engines installed on your own machine. The core generation loop in the scenario skill is broken into six steps: discover, schema, run, wait, display, and download.

What Problem It Solves

It tackles the old headache of getting an AI agent to reliably produce creative assets you can actually use. In the past, asking AI for images or video often returned things that could not go straight into a project: characters looked different from shot to shot, and the style drifted.

This skill set lays out concrete paths. Starting from a blank brief, it can produce reference images and mood boards. It keeps a single character or a single style consistent across a whole series. It reads generated assets back and runs batch review against the brief with a quality gate. It also derives the distribution sizes each platform needs from one master file. On top of that, it distills the operations of ZBrush, Blender, Maya, Unreal, Unity, Godot, and Roblox Studio into scripts an agent can drive directly.

If the Agent Skills format is new to you, read What Is mattpocock/skills and How to Use It first to get the basics in place.

Just as important, it has cost control built in so you do not burn money blindly. Before spending anything, it runs a dry run for pricing and then a quality gate to decide whether the job is worth placing.

Language and Architecture

  • Main language is Python. Script tests need Python 3.11 or newer plus ffmpeg. Skill definitions live as SKILL.md files, the CLI is distributed to 70 plus agents through the skills command, and the directory is skills.sh.
  • Scenario MCP integration: run claude mcp add --transport http scenario https://mcp.scenario.com/mcp. You can also add that URL in any MCP capable client and log in with a Scenario account (OAuth) or an API key.
  • Expert tool families, each with one lead plus several specialists, all verified in practice. ZBrush 2026 is 1+8, Blender 5.2 is 1+14, Maya 2027 is 1+10, Unreal Engine 5.8 is 1+9, Unity 6.3 LTS is 1+13, Godot 4.7 is 1+13, and Roblox Studio 0.743 is 1+13.
  • Model families covered: 8 image models (GPT Image, MAI Image, Seedream, Gemini Image, Reve, Ideogram, Grok Imagine Image, Luma Image), 10 video, 4 audio, and 4 3D. Each family has its own skill doc that guides routing between members and tiers.
  • Cost and governance come from three skills: scenario-quality-gate (free in stock reads plus dry run pricing), scenario-model-comparison (dry run cost matrix, billed cost, and latency), and scenario-team-admin (Creative Unit caps and spend attribution per member).
  • Testing: pnpm test runs Python unittest and TS vitest; pnpm run validate runs commitlint, cspell, prettier, and husky content checks.

License

MIT

In 2026 the route of agent skills plus generative AI for creative production is hot at the same time. This project packs production grade creative workflows straight into agents, combining 8 image, 10 video, 4 audio, and 4 3D model families with cost guardrails, and it can drive 7 families of DCC and engine tooling locally. Scenario itself is a recognizable creative AI brand. It gained about 667 stars that day and topped the new projects list.

Where to Find It

https://github.com/scenario-labs/skills (directory home https://www.skills.sh/scenario-labs/skills)

Getting Started

First confirm locally that your agent framework supports MCP, then wire in claude mcp add --transport http scenario https://mcp.scenario.com/mcp and log in with a Scenario account or API key. On the first run, start with the core scenario skill and walk the discover, schema, run, wait, display, download loop by generating one simple image, so you know the chain works before taking on complex tasks.

If you want the expert tools, install the matching DCC software on your machine first, such as Blender or Unreal Engine, and check that the version matches what the skill doc states. In production, run scenario-quality-gate for a dry run price and quality gate before you actually place an order, so the budget stays under control.

Summary

scenario-labs/skills pulls AI creative content work back from scattered prompt experiments into a skill set with a process, a quality gate, and cost guardrails. For teams that want agents to steadily produce game and entertainment assets, it removes the pain of building a generation pipeline and writing driver scripts yourself, and it is a ready made option worth wiring in.

FAQ

Do these skills need to be online to work? Scenario skills go through the cloud MCP server, so generating content needs a network call. The expert tools drive the software on your own machine, so they mostly depend on the local environment, but the licensing and dependencies of Maya or Unreal still have to be in place.

With so many models, how do I pick? Every model family has its own skill doc that explains routing between members and tiers. Newcomers can start with scenario-model-comparison, which gives a dry run cost matrix, billed cost, and latency to help you choose by budget and speed.

How does cost control actually work? It rests on scenario-quality-gate and scenario-team-admin. The first does in stock reads and dry run pricing before you spend, the second manages Creative Unit caps and spend attribution per member. Run a dry run before a real job and you get a fair idea of what it will cost.