Remotion Ships Skills 2.0 so AI Agents Write Video Code Smarter

Remotion just shipped a major overhaul of the way AI coding agents write video code. The framework, which turns React components into rendered MP4s, has retooled its Agent Skills system after admitting that the original markup was hard for both humans and models to follow. The result is Skills 2.0, a leaner, modular set of instructions designed to plug into Claude Code, Codex, Cursor, and Kimi Code without polluting the model's context window.

Why the rewrite was needed

Remotion Agent Skills are essentially rule files that teach a coding agent how to write correct Remotion code, covering compositions, sequences, timing, captions, and rendering. Remotion works well with coding agents including Claude Code, Codex, Kimi Code and OpenCode, but the previous skill pack was bloated and opinionated. Loading it into a session meant every generation had to fight against baked-in design defaults, and the sheer size of the rules ate into the agent's usable context.

The 2.0 rewrite fixes that with three headline changes: sub-skills, interactive output, and what the team calls a taste-neutral approach. Design instructions have been stripped out so the agent leans on the user's brief instead of clashing with hardcoded defaults.

How the new routing works

Instead of one monolithic skill, /remotion-best-practices now behaves as a router that dispatches to whichever specialized sub-skill is relevant to the prompt. That keeps token usage low because only the pieces that matter for the current task get loaded. You can also call a specific skill directly if you already know what you need. The available sub-skills cover most of the surface area of the framework:

• /remotion-create for scaffolding a new project or composition
• /remotion-markup for animations, layout, typography, audio, and timing
• /remotion-studio to launch the preview UI
• /remotion-render to invoke a render into a video or still
• /remotion-maps for static maps, animated routes and markers, geographic explainers, Mapbox, MapLibre, MapTiler, GeoJSON, and 3D geographic flyovers with CesiumJS
• /remotion-captions for subtitles and caption workflows
• /remotion-saas for architecture guidance when embedding Remotion in a product
• /remotion-interactivity to make generated elements selectable and editable in Studio
• /remotion-docs to fetch any documentation page as Markdown at generation time
• /remotion-upgrade to bump Remotion and related packages safely
• /remotion-multimedia for browser-based media handling with Mediabunny

The interactivity hook

The most interesting piece for anyone iterating on real videos is that skills now emit code you can hand-edit inside Remotion Studio. Generated elements are made selectable so you can nudge a timing, swap a color, or reposition an element in the preview, then hand the file back to the agent for the next round of changes. That closes a painful loop where agent output was either accepted wholesale or rewritten from scratch.

Installing and using it

Setup is a single command. From an existing project you run npx skills add remotion-dev/skills, or you can accept the prompt when running bun create video on a new project. There are also first-party plugins for each supported agent, listed on the plugins page, including the Cursor plugin distributed through the Cursor Marketplace.

npx create-video --yes --blank my-video
cd my-video
npm install
npx remotion skills add
npm run dev

What it is good and bad at

The framework itself shines at deterministic, brand-controlled motion graphics: logo animations, kinetic typography, product demos, data visualizations driven by JSON, and social clips with word-synced captions. Because output is React code rendered to video, results are reproducible with exact text, exact colors, exact timing, re-renderable forever. That is a very different tradeoff from diffusion-based generators like Sora or Runway, which produce pixels but cannot guarantee a specific string of text or a brand color.

The weakness is that quality still tracks how well you can describe motion. As one practitioner put it while testing the system, Remotion is never the bottleneck , motion-design craft is. Agents without a strong prompt will happily emit flat linear fades. The taste-neutral direction of 2.0 amplifies this: with defaults removed, vague prompts get vague videos.

Why it matters

The broader shift is that programmatic video is becoming a prompt-first workflow. Skills flip the entry point from thinking in components, frames, and composition from day one, which was fine if you already lived in code but created a high starting cost for anyone else. With Remotion Skills, the starting point shifts. Instead of asking how do I build this animation, you can start with what should this video communicate. For teams already generating marketing videos, explainers, or automated data recaps, 2.0 makes the agent loop cheap enough to iterate on hundreds of variants, then hand-tune the winners in Studio.