How to use Remotion agent skills with Claude Code
Installation is already well covered in Remotion's own documentation and in existing tutorials, so this page focuses on the outcome: a real edit on real footage with Remotion agent skills and Claude Code. It shows the prompts, what the run cost, what went wrong and the versions involved, so you know what to expect.
This page is built but kept out of search until it carries the evidence it needs: Needs recorded footage and a run with Remotion's skills; in testing.
What we ran and what we did not
Not run by us: the Claude Code step. The workflow below has an agent write and edit a Remotion project from a set of skills. Running it needs an agent account, and we do not create accounts for tests. So this page does not report how well an agent does the job, how long it takes or what it gets wrong. Anything you read about that step on this page is a description of the plan, not a result.
What we did run: Remotion itself, the renderer those skills drive, on the same Tears of Steel clips we use for every test. Two tests are published: We tested Remotion's TikTok captions template on three clips (the TikTok captions template, as shipped) and We tested Remotion's transitions package on three clip pairs (fade, slide and wipe on three clip pairs).
The workflow in short
- Write the target: length, aspect ratio and style.
- Install Remotion's agent skills in your coding agent and give it the brief and the footage.
- Let it build a Remotion project. This is the step we have not run.
- Render it with Remotion's command line and review the result against the criteria below.
- Record the fixes a person still has to make.
What we measured in Remotion itself
The captions template transcribes with whisper.cpp and renders a 1080x1920 video. On this CPU-only laptop its transcription took a median of 453, 204, 126 seconds for our three 60-second clips, and the render took a median of 217, 229, 215 seconds. Peak memory reached 5.8 GB.
The transitions package joined 6-second clip pairs with a fade, a slide and a wipe. Each 11-second render took a median of 4 to 43 seconds, and every transition moved from the first clip to the second as expected.
Whatever an agent writes, it still ends in these renders. So these are the render costs and limits you should plan around. Read the failure boxes on the two test pages before you hand the job to an agent.
Acceptance criteria
- The export meets the target length and aspect ratio.
- No cut lands mid-word, and no wanted pause is lost.
- Captions are correct and inside the safe zone.
- Audio is clean and even.
- A person watched the whole export, and that time is counted.
Method and limits
- The agent step was not run, so this page makes no claim about it.
- The Remotion numbers come from our two tests on three 60-second clips of one film, on a laptop CPU shared with other work.
- Check Remotion's licence for your use before you build on it: it is free for some uses and paid for others.
- Tool versions are on the test pages, and results can change when they update. Our method is on How we curate and test AI video workflows.
The run uses real footage and the tool versions stated on the page. We record the prompts, the cost, every failure and its fix, and we link to the sources of the skills and the documentation they rely on.
- Skills and tools change quickly, so every result states the versions it used.
- One run on one set of footage does not show every outcome.
- Check the Remotion licence for your intended use.
Page checked on Nov 4, 2026.