Why Claude Remotion Videos Look Flat and How to Fix It
AI Tools Analysis
Why Claude and GPT-6 Astra Remotion Videos Look Flat (and How to Fix It)
October 7, 2026
Keston CollinsVideo editor with nearly 10 years of experience, exploring the intersection of motion graphics and AI.
Videos made with Claude Opus 5.5 or GPT-6 Astra look flat for a simple reason: neither model draws video. Both write Remotion code, which is React, and Remotion renders that code into frames. Anything the code can't draw with text, shapes, gradients and images never shows up on screen. That's why so many of these clips read like a slide deck that learned to move. You can push the result further with better prompts, real fonts and real assets, and for the few seconds that need to look designed, you can drop in a finished clip instead of asking the model to invent one.
What Claude and Astra are actually doing
When you ask Claude Code to "make a launch video," Opus 5.5 writes a set of React components: one per scene, each one reading the current frame number and moving things around with spring() and interpolate(). Remotion then opens those components in a headless browser and records them frame by frame into an MP4.
GPT-6 Astra works the same way underneath. Remotion maintains a plugin for Codex in the ChatGPT desktop app (Plugins tab, search "Remotion"), and once it's installed you type $remotion in a new project and describe the video. Astra's strength is running the whole job on your computer: one popular walkthrough by Rob Cressy has it cutting raw footage, adding captions and graphics, and rendering the final file in a single session. The graphics part is still Remotion code.
So the visual ceiling is the same for both models. It's whatever a web page can render. Text, rectangles, SVG paths, CSS gradients, blur and images are all available. Lighting, materials, particles that feel physical, and real 3D are not, unless you bring them in as assets.
Three reasons the output looks flat
Everything is drawn with the same few tools: dark background, glowing headline, soft gradient blob, pill-shaped buttons. The model reaches for what is easy to express in CSS, so most first drafts land on the same look.
Motion defaults repeat. Unless you say otherwise, every element fades and slides in with a similar spring, so scene two feels like scene one.
Fonts and assets fall back. If a font isn't loaded properly, the render falls back to a system font. If you don't provide screenshots or footage, the model fills the frame with type, because type is the only content it has.
What a side-by-side test showed
To see how far prompting gets you, the same brief was run two ways on October 7, 2026. The brief: a 16:9 launch opener (brand wordmark, the line "Motion graphics, ready to post," a three-step subline, brand purple and orange) and a 9:16 Reels hook about not rebuilding motion graphics from scratch.
Route one was Claude Opus 5.5 writing Remotion from an empty project, then one round of the most common feedback ("it's too static, add depth"). Route two was a designed template from AutoAE's library, with only the text and five product screenshots swapped in.
Claude Opus 5.5 + Remotion, after the second round: grid floor, light sweep and a slow push-in were added, but the frame is still type on a dark gradient.
Template route: particle logo reveal, product screens orbiting the core line, a light-beam background. One render, no code.
Launch opener
Claude Opus 5.5 + Remotion
Designed template (AutoAE)
Time to a designed clip
About 2.5 minutes for two rounds, and still not at a designed look
About 1 minute plus uploading screenshots, designed on the first render
What you write
186 lines of React by the second round
Two lines of text, five screenshots
What the frame contains
Type, gradients, a CSS grid, a light sweep
Particles, orbiting product screens, light beams
What it costs
Model usage for every round you write and rewrite; Remotion itself is free for individuals and small teams
$12 a month on Starter covers about 50 downloads (about 24 cents a clip); previews are free
The Reels hook told the same story. Claude's first version broke "motion graphics" across two lines and needed a second round to fix the layout. The template version opened on a 3D character and moved through three scenes in about five seconds.
Claude Opus 5.5 + Remotion, second round.
Template route, one render.
In both tests the template route reached a designed clip on the first render, while the code route took two rounds and still wasn't there. The template was also faster on the clock: the launch opener took about a minute plus uploading screenshots, against about two and a half minutes for Claude's two rounds. After those rounds the Claude version was cleaner, not more designed. Nothing in the feedback loop gave it materials, lighting or a camera, so getting there means building each of those by hand, starting from an empty file, and paying for model usage every round. With a template library the motion design is already done: you pick a template, type your text, and render. This was a single brief, so treat it as one data point, not a benchmark.
How to make Claude or Astra Remotion videos look less flat
Write motion direction into the prompt. "Make it cool" gets defaults. Say what moves, in what order, and how: "headline words stagger in 4 frames apart, spring with damping 12, hold 1.5 seconds, exit upward while the background keeps drifting." Ask for one easing curve per element type so the scenes don't all feel the same.
Load real fonts. Use @remotion/google-fonts or load your brand font with loadFont before the composition renders, then check a frame. A fallback system font is the fastest way to look generic.
Give the model assets, not just words. Screenshots, screen recordings, product photos and pre-rendered 3D elements give the frame something besides type. Remotion plays them with <Img> and <OffthreadVideo>.
Build layers on purpose. Ask for a foreground, midground and background that move at different speeds, with blur on the back layer. Parallax is the cheapest depth you can get in code.
Check frames before you render the whole thing.npx remotion still renders a single frame. Ask Claude or Astra to export three or four stills at the key moments and look for overlapping text, broken lines and empty stretches where nothing moves. The broken line in the Reels test would have shown up in the first still.
Where a designed clip earns its place
Most of a video doesn't need to look expensive. Captions, lower thirds, simple title cards and data callouts are exactly what Remotion code handles well, and keeping them in code means you can change a word without re-exporting anything.
The moments that carry the first impression are different: the hook in the first two seconds, the logo reveal, the launch shot that shows the product. For those, a clip a motion designer has already built usually beats anything you can prompt in a few rounds. If you already work in Claude Code or Codex, AutoAE's Motion Graphic Agent can run inside that chat through the AutoAE Skill (install it with npx skills add https://autoae.online, Node.js 24 or later). Inside the same chat, the agent searches a library of professionally designed templates, fills one with your text and images, and returns an MP4. Drop that file into your Remotion project with <OffthreadVideo> and keep building the rest in code. The two clips above, the launch opener and the surreal 3D Reels intro, came from that library.
FAQ
Can Claude make videos by itself?
Not directly. Claude writes code, and a tool like Remotion turns that code into video. In Claude Code you can install Remotion's Agent Skills so Opus 5.5 knows the API, then render locally with npx remotion render.
Is Remotion free to use?
For individuals, for-profit companies with up to three employees, and non-profits, yes, including commercial videos. Larger for-profit companies need a Remotion Company License. Check the current terms in Remotion's license before you ship client work.
Does GPT-6 Astra make better videos than Claude?
Both write Remotion code, so the visual ceiling is the same. Astra is built for running long tasks on your computer, which helps when the job includes cutting real footage and captions. For the graphics themselves, the prompt, the fonts and the assets you provide matter more than which model writes the code.
