Motion Graphics Just Had Its Claude Code Moment

We build an AI creative platform that runs on the best generative models on the market. And our own feature videos were still the hardest thing we shipped.

Sounds like a contradiction. It isn’t. Kreator can spit out a product photo, a UGC-style clip or a claymation scene in minutes. What it can’t do on its own is finish a video. Generated clips top out at a few seconds. They have no clue what our host is about to say at 0:31. They mangle on-screen text. And they don’t know our orange is #FF704D unless somebody pastes it into the prompt, or saves it as a skill or system prompt.

So every walkthrough, webinar welcome and launch clip died the same death. We had a decent recording and a folder of great assets, and then came a long, mechanical edit nobody on the team had time for. We either shipped it half-baked or let it sit until it went stale.

Same take, same second. Left: what Kai recorded. Right: what came out of one conversation.

What We Tried

Kai on our team recorded the opener for our 15-minute core features walkthrough. One take, talking to camera in a natural setting. That’s it.

The raw take, 52 seconds.

Instead of opening After Effects or chasing a contractor on Behance, I opened Claude, wired into three things:

• Kreator, through our MCP server. Claude could look up products, pull existing assets and generate new images, sound effects, a music bed or clips. Every generation came back as a dry-run quote first, so nothing burned credits without my OK.

• Our brand guidelines, pulled straight from Notion: navy and orange, Gelasio italic for emphasis, Inter for everything else, and how we talk about the product.

• HyperFrames, an open-source tool that turns HTML and animation code into MP4. It lives in a small motion folder on my Mac, next to our brand tokens and a handful of reusable animation helpers.

Then I briefed it the way an executive briefs a motion designer. In other words, badly:

Here is a new video that we want to use as an intro video for a 15-minute feature overview. Make edits and tweaks, add some enhancing visuals. Stick to brand guidelines as much as possible.

Claude turned that into an actual creative brief:

Keep Kai on camera, use the empty wall on the right, illustrate each thing she names as she names it, stay on brand, and hand off to the screen recording at the end.

What Came Out

The finished opener.

A 52-second opener where the empty side of the frame keeps up with Kai, beat by beat:

• She says “it all starts with your product URL,” and a card types out a URL with Website and Amazon options.

• She lists what Kreator pulls in (info, images, brand messaging, audience segments), and a product profile ticks off each row as she says it.

• She talks about consistency, and the profile becomes the foundation for a grid of on-brand ads, each tagged with the audience it targets. The brand is Brightfield Cold Brew, a fake company we built inside Kreator for exactly this kind of demo.

• She names UGC videos, product videos, captions, animation and claymation, and a deck of real examples fans out. Some were fresh generations. Others were tiles we’d already made for our tools page.

• “Let’s see it in action” wipes to navy, and our wordmark hands off to the screen recording.

Beat by beat: product URL, product profile, on-brand ads, captions, static ads, hand-off.

Bonus: a cleaned-up caption file (“Creator” fixed to Kreator, “acclaymation” fixed to claymation), dead air cut from both ends, and a music bed that ducks under her voice.

I never touched a timeline. I looked at stills, said things like “the labels flash by too fast” or “zoom in on Kai a bit less,” and got a new render. We made a couple more videos the same way: a launch clip for the Kreator MCP and two cuts of a welcome video for our AI Ick webinar. None of them needed a production schedule. Are they perfect? Absolutely not. Are they good enough, and better than not shipping at all? Absolutely yes.

Why It Works

The trick is splitting the job in two. Generative AI makes the pixels. Code makes the motion.

Generated video is great at stuff that’s hard to fake: a person holding a can on a hiking trail, a claymation scene. It’s terrible at precision: exact text, exact timing, a 52-second structure that follows a script. Code is the opposite. A Gelasio headline at 64px that lands on the word “brand” at 18.4 seconds is trivial in HTML and close to impossible to prompt for.

Give Claude the code, the generator and the brand guide, and the gaps close. Duration stops being a limit, because the video is as long as the composition. Text comes out right because it’s real type, not hallucinated pixels. And context comes through because Claude has read the transcript, the product profile and the guidelines before it draws a single frame.

Context is the real multiplier. The same request without it gets you a generic explainer that could belong to anyone. With it, you get our colors, our fonts, our demo product, our existing assets, and a host who doesn’t get buried under stock graphics.

What Wasn’t Magic

It wasn’t frictionless. The rough spots tell you where this is headed.

• Plumbing. The cloud sandbox Claude worked in couldn’t install the HyperFrames CLI or pull assets from our own storage. So we fell back to my laptop, and with it the classic “works on my machine” headaches.

• A bug that came back. My local render had the music way too loud. The renderer was multiplying two volume settings that were supposed to replace each other. Claude found it, fixed the renderer and wrote the rule into our motion README so it stays fixed.

• Not every edit is possible. Kai says “you know” around the 30-second mark. Without word-level audio timing we couldn’t cut it cleanly, so it stayed.

• Taste still needs a human. The first reframe of Kai was too tight, and some labels flashed for a fraction of a second. Claude checked its own frames and caught some of that, not all. The job moved from doing the edit to judging it.

The Reckoning

I’ve seen this movie before. It was called software.

For years, being a good developer meant knowing the framework, the language’s quirks and the IDE shortcuts. Coding agents made that knowledge cheap almost overnight. What’s left is knowing what to build, spotting when it’s wrong, and having the taste to make it good.

Motion graphics is next. The craft has always been locked behind tools: After Effects keyframes, Premiere timelines, plugin stacks, render queues. You needed years of fluency in one app’s UX and vocabulary before you could make anything useful. That moat is draining. The browser has been able to do serious motion since HTML5. The outline of this painting has been sketched for a decade, and agents are now putting on the final strokes.

For small teams like ours, this is a massive unlock. We don’t need a motion designer on payroll to ship polished, on-brand video every week. For motion design as a discipline, it’s an upgrade, not an ending. For a lot of individual motion designers, it’s going to hurt. The ones with the best eye just got a tool that turns that eye straight into output, with no hours of keyframing in between. The ones who’ve been a cog on the production line need a new area of expertise, and they need it fast.For small teams like ours

What’s Next for Us

Kai’s opener is episode one of a walkthrough series headed to YouTube and kreator.ai. Every video in it gets the same treatment, built from the same brand kit and helpers, and the kit gets sharper with every video.

We’re also pushing more of this into Kreator itself. The MCP server is what made this whole thing possible, and you can plug it into Claude, ChatGPT, Cursor or whatever agent you use today. Try it on your own recordings and send me what you make.

You still have to hit record. Everything after that is just a conversation.