Remotion: 49k Stars and 18 Agent Skills for Video in React
- Smars
- Agent Skills , Open Source
- 06 Jun, 2026
Write a video in React.
This is the premise of Remotion, and it has proven to be a surprisingly powerful idea. 49,200 stars, 3,500 forks, 633 releases, used by 4,700 projects. Fireship made a video with it. GitHub Unwrapped was built on it. Companies pay for licenses to use it in production.
But this article isn’t about what Remotion does. It’s about how Remotion maintains itself.
Scroll to the root of the Remotion monorepo and you’ll find a directory: .agents/skills/. Inside, 18 skill directories. Each one teaches an AI coding agent how to perform a specific task within the project.
The Skills Remotion Ships
Not as a plugin. Not as an add-on. These skills are part of the repository itself.
Building (add-cli-option, add-effect, add-new-package, add-sfx):
Step-by-step workflows for implementing new features. The add-effect skill covers shader compilation, schema definition, lifecycle hooks, bundle registrations, tests, docs, demo pages, preview images, and the order in which to do them. An agent reading this can add a visual effect to @remotion/effects without needing a human to explain the architecture.
Documentation (writing-docs, docs-demo):
Remotion’s docs are MDX. Each API gets its own page. Sidebars need updating. Preview cards need rendering. The writing-docs skill encodes the conventions an agent must follow — breadcrumb format, heading hierarchy, twoslash code snippets, the Demo and Steps components, and the exact command to run for preview card generation.
Project management (issue, issue-management, pr, pr-name, pr-ready):
PRs must follow a naming convention: [package-name]: commit-message. The pr-name skill enforces this. The pr-ready skill checks that all required files are modified, docs are updated, and tests pass before an agent marks a PR as ready.
Maintenance (version, update-version, fix-dependabot):
Version bumps, changelogs, and automated dependency fix PRs. The agent reads the current version from packages/core/src/version.ts, increments the patch, and updates all references.
Debugging (video-report, visual-mode, web-renderer-test):
When a rendered video breaks, the video-report skill tells the agent how to download the failing file, set it as the source in the testbed, and run a verbose render to capture logs.
Why This Exists
AGENTS.md files are fine for project-level instructions. But they scale poorly. A single AGENTS.md for a 49k-star monorepo with dozens of packages, a custom license model, and a Rust-powered rendering engine would be a 10,000-line document that an agent can’t effectively use.
Remotion’s approach is skill decomposition. Instead of one enormous instruction file, the project has 18 focused skills. An agent working on an issue about a broken effect doesn’t need to know about the PR naming convention or the docs sidebar structure. It reads add-effect/SKILL.md and gets exactly what it needs.
This matters because the bottleneck for AI contributions isn’t capability — it’s context. A capable agent given the wrong context fails. A capable agent given precisely the right skill file succeeds.
What Makes Remotion’s Skills Different
The skills we’ve covered in this series so far — edulab, Seedance2-Skill, Videocut Skills — are all skills users install to extend their agents. They’re consumable tools.
Remotion’s skills are different. They’re embedded in the project source. They exist not for users of the software, but for contributors to the software. They’re a form of executable documentation — instructions written not as a README for humans to follow manually, but as a skill file for agents to read and execute.
This is a new pattern. A project ships its contribution guide as agent-readable skills alongside the source code. New contributors — even AI contributors — clone the repo and have everything they need to start making meaningful changes.
The Skills at Work
A developer files an issue: “The @remotion/effects package could use a pixelate transition.”
A maintainer or an AI agent assigns themselves. The agent reads add-effect/SKILL.md:
- Check if the effect shape already exists in the registry
- Create the WebGL2 shader file with vertex and fragment programs
- Implement the schema with
SequenceSchema— duration, params, defaults - Wire up
setup(create WebGL program),apply(bind uniforms, render),cleanup(delete resources) - Export from the package bundle
- Add tests for default params, edge cases, and key distinctness
- Write the doc page with a
Democomponent and a preview composition - Render the preview card
- Format, build, verify
Each step is documented with code templates, file paths, and commands. The agent doesn’t guess. It follows the recipe.
Try It
Remotion is at github.com/remotion-dev/remotion. The skills live in .agents/skills/. Clone it and see what an agent can do with 18 focused instruction files.