A Video Studio Inside Your Agent
Programmatic video: Claude writes React components, Remotion renders to MP4 — no timeline editor. A heavy kit adds voiceovers, captions and a render-review-fix loop.
AI video has the same slop problem as AI images, just moving. Drop a prompt into a generic video tool and you get a template everyone else got too — the same kinetic-text intro, the same stock B-roll, the same library music — with your logo dropped in. It's the median, set to music. And because it's a flat rendered file, you can't fix the one wrong word without re-rolling the whole thing and hoping the rest survives.
There's a different model: programmatic video. You don't open a timeline editor — Claude writes React components, and Remotion renders them to MP4. Your video becomes code: diffable, parameterizable, version-controlled, and editable by changing a value instead of dragging a clip. This is Layer 2 anti-slop at its most ambitious — instead of generating a flat asset, you're building a real, programmatic one. The studio lives inside the agent.
Key Takeaways
- Video as code, not as a timeline. Claude writes React components and Remotion renders them to MP4 — no editor, fully diffable and parameterized.
- Editable, not baked. Wrong word on screen? Change a string and re-render — you don't re-roll a whole generated clip and pray.
- The heavy kit is a studio. remotion-superpowers-video-studio adds voiceovers, music, captions, and a render → review → fix loop.
- The light kit is the on-ramp. remotion-kickstart-create-video scaffolds a project with one command — the official way in.
- Pick by ambition. Light kit for a quick parametric clip; heavy kit when you need narration, captions, and an automated review loop.
Why Programmatic Video Beats Generated Clips
A generated video clip is a flat file: opaque, unfixable, and identical to whatever everyone else generated from the same template. Programmatic video treats the video as a program — and programs have properties flat files don't.
| Property | Generated video clip | Programmatic video (Remotion) |
|---|---|---|
| Source | Opaque rendered file | React components — readable, diffable |
| Editing | Re-roll the whole clip | Change a value, re-render |
| Text on screen | Baked pixels, can hallucinate | Real text, copy-paste accurate |
| Reuse | One-off | Parameterize once, render many |
| Version control | Binary blob | Plain code in git |
| Consistency | Drifts every regeneration | Deterministic from the same source |
| The "tell" | Template + logo = median | Built to spec, ownable |
The "editing" and "consistency" rows are the unlock. A typo in a generated clip is unfixable without re-rolling everything; a typo in a Remotion comp is a one-character change and a re-render. And because the same component with different props renders deterministically, you can produce ten variations of a video the way you'd produce ten variations of a chart — by passing different data.
The Light Kit: Your On-Ramp
If you've never touched Remotion, start light. remotion-kickstart-create-video is the official on-ramp — it scaffolds a working video project so the agent has something real to edit from line one.
npx create-video@latest
That gives you a React project where compositions are components and props are your parameters. From there the agent works in plain React:
"In this Remotion project, build a 15-second intro composition.
Animate the brand wordmark in with a spring, then a one-line tagline
fades up beneath it. Use brand primary #FF6B35 on near-black.
Expose the tagline as a prop so I can render variants. Render to MP4."
Change the prop, re-render, get a new variant. No timeline, no dragging, no re-rolling — just code that produces video. This is the right entry point when you want a clean, parametric clip without the full studio machinery.
The Heavy Kit: A Full Studio Loop
When you need more than a clip — narration, music, captions, and a way to catch mistakes — reach for the heavy kit. remotion-superpowers-video-studio wraps Remotion in a production pipeline: it adds voiceovers, background music, captions, and crucially a render → review → fix loop so the agent inspects its own output and corrects it instead of handing you a broken render.
/plugin marketplace add DojoCodingLabs/remotion-superpowers
/plugin install remotion-superpowers@remotion-superpowers
/setup
Run /setup and you've got the studio. Now the brief can be a whole short video, not a single comp:
"Produce a 45-second explainer on killing AI design slop. Script the
voiceover, add subtle background music, and burn in captions synced to
the narration. Brand palette primary #FF6B35 on near-black, DM Sans.
Then run the render → review → fix loop: render, check timing and
caption sync, fix any drift, re-render the final MP4."
Choosing between the two is just a question of ambition.
| Light kit — remotion-kickstart | Heavy kit — remotion-superpowers | |
|---|---|---|
| Setup | npx create-video@latest | plugin install + /setup |
| Best for | A quick parametric clip | A produced short video |
| Voiceover / music | Bring your own | Built in |
| Captions | Manual | Generated and synced |
| Review loop | You review | Automated render → review → fix |
| Mental model | "A React component that renders to MP4" | "A studio that ships a finished video" |
Both share the same anti-slop core: the video is code, so it's yours, it's editable, and it's never just a template with your logo dropped on top.
Frequently Asked Questions
Do I need video editing experience to use Remotion?
No. You never open a timeline editor — the agent writes React components and Remotion renders them to MP4. If you can describe the motion you want and read a diff, you can ship programmatic video. Start with remotion-kickstart-create-video, which scaffolds a working project so the agent has something real to build from.
What's the difference between the heavy and light kits?
The light kit (remotion-kickstart-create-video) is the on-ramp: one command scaffolds a project for quick, parametric clips. The heavy kit (remotion-superpowers-video-studio) is a full studio — it adds voiceovers, music, captions, and an automated render → review → fix loop. Use light for a clip, heavy for a produced short video.
Can I fix one wrong word without re-rendering the whole thing?
You re-render, but you don't re-roll. Because the video is a React component with real text, fixing a word is a one-character change, and the re-render is deterministic — everything else comes back identical. That's the opposite of a generated clip, where any fix risks a completely different output.
How do I make ten variations of the same video?
Parameterize it. Expose the changing parts (a tagline, a name, a stat) as props, then render the same composition with different prop values. You get ten on-brand variants from one source the way you'd get ten charts from one template — by passing different data, not by rebuilding each one.
Browse 135+ agent-ready design systems in the Designs category, or explore the full skill catalog at aiskill.market.