Klea Merkuri

Klea Merkuri

Nov 19, 2026 · 14 min read

Can AI Make a Polished Animation From One Prompt?

Motion demos made with Opus and Fable are everywhere right now, usually in the form of a clean (rather cool) video with a line or two of prompt in the caption or comments. The unspoken message is that you typed that, and the model did the rest.

I kinda believed it, enough to test it on my beagle, Eve. That one’s on me 🙈

I gave three Claude models a photo of Eve in a birthday hat and asked each one to turn it into a 7-second loop.

The idea was that she assembles out of pixels starting from her face, holds there alive for a few seconds with a few sparkles rising from the candles on her hat, then falls away like sand with her eyes and nose leaving last.

Her face had to stay exactly her face.

An animation of one dog in one hat seemed like the easiest possible version of the claim. If it was going to work anywhere, it would work here.

What came back was good but nowhere near what the demos advertise.

Can one prompt do it? All three got close, and none of them finished.

What Does a Good One-Shot Prompt for an AI Animation Look Like?

This is the whole prompt, word for word, and the only thing any of the three models saw. “One-shot” means exactly that: one prompt and no follow-ups.

Goal: turn the photo eve.jpg in this folder into a seamless looping animation, about 7 seconds, that makes my dog Eve appear to come alive out of pixels and then fall away again. This is a capability test at medium effort and the finished loop is for posting, so what matters is what a viewer sees and feels — her likeness has to survive it. How you build it is your call.


What the viewer sees, in order:

1. She assembles out of pixels, starting from her face (centre of the frame, a little above the middle) and spreading outward — face first, then her birthday hat and body, background last. She also sharpens as she assembles: blocky and abstract at first, ending as the full sharp photo.

2. She holds, alive, for roughly 3 seconds. The motion stays subtle — a gentle sense of breathing and presence, a few small pixel sparkles rising from the candles on her hat. Her face must stay exactly her face: no redrawing, warping, or generating features.

3. She dissolves away fast, in about 1.5 seconds, as falling pixels — like sand dropping out of the picture — with the background going first and her eyes and nose leaving last. Then the loop restarts with her eyes returning first, with no blank or black frame in between.


Deliverable: one self-contained HTML file that just works when opened in a browser, with the photo loaded from the same folder. No network access or external services needed to run it. It should hold a smooth frame rate on an ordinary laptop, keep the photo's portrait orientation, and fill the viewport sensibly. Include a play/pause control and a way to export one full loop as a downloadable video file. No text overlays, no music, no 3D.


Before you report done: actually run it, capture a still from each of the three moments (assembling, alive, dissolving), look at them, and fix anything that is broken, distorted, or janky before reporting. Report the file, what you checked, and anything you had to compromise on.

Take note of that last paragraph, because I didn’t only ask for the animation.

I asked each model to run it, capture a still from each of the three moments, look at them, and fix what was broken before reporting back.

All three models did exactly that.

Why Did I Rewrite the Prompt Before Anyone Saw It?

Now, the first version of this brief was a step-by-step list. It told the model which canvas to use, the exact tile sizes, the exact zoom percentages, how many sparkles, and which export method.

I rewrote it before any model saw it, going by Anthropic’s current guidance for these models.

Give them the end result and the boundaries, and they plan the steps better than a prescribed sequence does.

What I kept were the calls a model can’t guess:

  • she assembles face first and background last
  • her eyes and nose leave last and come back first
  • the loop restarts with no blank frame

The hold runs about 3 seconds, the dissolve about 1.5, and her features can’t be redrawn, warped, or generated.

Tip 👀
That rewrite is now my standing rule for how I prompt Claude models. Each one gets the goal and why, the end result as someone experiences it, context the model can’t see, boundaries only, and done-evidence.

Related: Is Prompt Engineering Still Worth Learning? What to Learn Next (and When)

Why Did I Test Claude Models Without Skills, brag, or HyperFrames?

Before any run, I decided to leave out skills, brag, HyperFrames, and packages. A skill is an add-on set of instructions you install so an agent knows how to do one kind of job. brag is one of those skills, a director skill for multi-scene promo videos.

My brief already held all the direction, plus a one-photo loop has no scenes or copy for it to direct.

HyperFrames is the video framework underneath brag, and I left it out for two reasons. First, it would have muddied the verdict on the models.

Second, it renders timelines you can jump around frame by frame, while the effect I’d asked for is sand-and-tile physics where each frame builds on the one before, so using it would have meant rewriting the effect before I’d even tested the model.

I wanted a clean verdict on the models 💁‍♀️

In hindsight, it’s also the cleanest way I could have found to see what the demos hide, because the layer they hide is the exact layer I’d taken out.

What Happened When I Gave Three Claude Models the Same Prompt?

Fable 5.2, Sonnet 5.5, and Opus 5.5 each got that prompt word for word, in a fresh session, at medium effort.

Medium suits interactive work and is where these models are more likely to stop early, so the question was whether any of them could do this on medium.

Now, I’m not ranking them. One run each isn’t enough to rank anything, and the differences between them are small enough to be noise.

I also kept my opinion out of it until the exported videos were pulled apart frame by frame at matching timestamps, so it couldn’t tilt the analysis.

Note: A follow-up like “fix the ending” would have turned a one-shot test into a conversation, so I didn’t send any. The entire point was to see what can a single prompt get you and how far can it get you for each of the models, not what can be achieved across multiple iteration rounds.

Why Did Sonnet 5.5’s Animation Surprise Me?

Sonnet was the surprise and finished in the middle. Watching it run, the animation was more graceful than I expected, and kind of complete, which is the awkward hedge I’m stuck with because it still wasn’t what I had pictured or what I thought I’d said in the prompt.

Pulling the frames apart reveals that around the 5.9-second mark the background has fallen away and Eve stays completely solid, which is the graceful part.

But the start is different from what I asked. At 0.3 seconds, the frame is confetti, with scattered pixels across her face, hat, and shirt at once, where the prompt said face first. It goes from scattered to sharp more than from chunky to sharp; the background fills in last behind a black crackle pattern, and the dissolve scatters instead of falling like sand.

And the file is where it comes up short. The export, meaning the video file it rendered, runs 6.29 seconds against the 7-second loop I asked for. Its last frames, at about 6.25 seconds, show the whole image mid-scatter, and that’s where the video stops, with no finished loop and no match between its first and last frames.

Sonnet’s summary is detailed and accurate about the page it built. The loop seam is pixel-identical, with zero differing pixels, and playback held around 115 frames per second.

For the export, it hit the button once and got what it described as a roughly 7-second MP4. Though the page loops, the video I got certainly doesn’t.

The summary warns that export records in real time and drops frames if the laptop stutters. So, I can’t tell how much of the missing 0.71 seconds is the model and how much is the recorder that day.

How Close Did Fable 5.2 Get to the Effect I Wanted?

Fable was good, not great, which is what I expected, so there’s no surprise in either direction. It came close to the whole effect, even though it didn’t fully finish.

Its opening is the one I’d pictured. At 0.3 seconds, the frame is only her face, as coarse blocks on black. The sparkles are tiny white and yellow plus signs.

The dissolve goes a step further than I asked. I wanted her eyes and nose to leave last, and at about 6.4 seconds her whole face is still sharp while her body drops away as a column of falling sand.

The part that didn’t match was the file. Fable’s summary says the export would be 1080×1440, because the page caps the working size at 1440 pixels tall. It also plainly states that it hadn’t re-measured the frame rate or the export at the new size. The export came out at 996×1328.

At the end, it plays fine and runs 7.02 seconds. The summary said what hadn’t been checked, and the export is where the difference showed up.

What Did Opus 5.5 Do Differently, and Did It Help?

Opus took way too long. It was also the only one that actually opened a browser to check what it was doing, and its summary reads like a running log of someone fixing things as they find them.

It built the effect in WebGL, captured stills from every phase, and compared the held frame to the source photo, which matched almost exactly.

Then it found problems 😬

The restart was nearly black, so it kept Eve’s eyes and nose visible through it. Its first export froze for about a fifth of a second at the start and ran about 0.3 seconds long, so it fixed that. The final export is exactly 7.00 seconds at 1080×1440, which is what its summary said, because it tested the export itself and not only the page.

Tip: Check the exported file’s length and resolution yourself, since a model’s summary covers what it built and the file can differ.

Frame by frame, the face leads, but noisily. At 0.3 seconds, the hat and body are already pixel dust around it. The background resolves through heavy black pixel-noise bands, and the dissolve crumbles to dust more than it falls like sand, though her eyes and nose do hang on as a ghostly face at the end.

The sparkles are where it went too far.

I asked for a few small pixel sparkles, but Opus found them too small and too faint against the bright sky. It made them bigger and more numerous, and the fix overshot into multicolor pixel clusters at the candles, closer to arcade confetti than candle sparkle.

I asked for small sparkles and got a small party.

How Complete Were the Results From One Prompt?

The prompt asked for done-evidence, and all three delivered it, with stills from the three moments, a list of what they checked, and a section on compromises.

Those compromises are also candid. Sonnet pointed out that the candle tips sit at the very top edge of the photo, so its sparkles leave the frame early. Opus admitted that Eve’s outline, eyes, nose, and candle tips are placed by hand for this photo, not detected automatically, so a different photo would need them placed again.

All three also kept her face intact, which was the one hard boundary in the brief.

One thing none of them were told is what should be behind the pixels while she’s partly assembled. All three chose a black void on their own.

Whatever you leave unspecified, you get the same default from every model.

Complete, for me, means Eve assembling out of pixels, coming alive, and dissolving in a clean 7-second loop, with sparkles that stay small. None of the three delivered all of that.

Maybe any of them could get there eventually, with a second round or a third. They definitely didn’t get there in a single prompt, and this prompt was already written for these models, since it gave the goal and the intent instead of steps.

That’s what annoys me. All three are fairly good at this, and not one of them fully completed everything correctly.

What Do AI Animation Demos Leave Out of the Prompt Caption?

A caption has room for a prompt. It doesn’t have room for the setup, add-ons, retries, or a person watching the output. And many of those prompts are one or two lines, which gives the wrong idea of what the model can do on its own.

Then you try it yourself, and yours doesn’t come out like the video.

My guess, and it is a guess, is that the original took multiple rounds, plus plugins, connectors, or packages, some paid and some free. I can’t see inside any particular demo, nor am I pointing at anyone.

What I’m pushing on is the format that feeds a hype devoid of details.

The reality of even the most capable model’s cool video lies in the workflows behind it. From the packages and connectors to the rounds of iteration and the person who watches each output, asks again, and picks the take worth posting.

That person is also deciding which tools fit the job, and a tool that fits one effect can be wrong for another.

Each of those layers changes what the agent starts with. A skill gives it instructions for one kind of job before it writes a line, and a package or connector gives it tools like a renderer.

A round of iteration adds a critique of its own first attempt. My three bare models got none of it.

Are the Models the Problem, or Is the Workflow?

I’m fairly convinced the model isn’t where the gap is. A bare model can do a lot now.

With one prompt and no tooling, all three models produced a working animation of a beagle assembling out of pixels, breathing, and dissolving, with her face untouched.

None of this is a knock on them.

Letting the model check its own work doesn’t close the gap either. Opus did exactly that, in a browser. It took the longest of the three to get there, and “there” wasn’t too far. Looking at its own output made the animation better and still left it incomplete.

Sonnet was the fastest, and its file is the only one that came up short. Opus was the slowest, and its export is the only one that matched its summary.

One run each doesn’t make that a rule, but checking costs time, and a caption never shows how long someone waited.

What’s left sits around the model, in the instructions that teach it a specific toolchain, the rounds where something gets corrected, and someone deciding when it’s done.

Though I can’t prove that’s what’s behind any particular demo, I can tell you what a bear model does because I ran it and that every one of those pieces exists as installable tooling.

The gap between my three videos and the polished ones online is mostly that layer, and the model is the smallest part of it.

Related: How to Get Started With Claude Code if You’re Coming From Cursor

It’s a Wrap

What will you ask next time an AI demo impresses you? I know that next time a demo impresses me, I’m skipping “what prompt did they use?” and asking what workflow they’re not showing.

If you’ve tried to copy one of these and yours looked nothing like the video, you were missing a production process nobody put in the caption.

Before you follow or recreate a demo, ask what’s installed and how many rounds happened between the prompt and the post.

And for anyone who wants to know what a model does on its own, run it bare once, with a detailed brief, one prompt, no add-ons, and a done-evidence line at the end, then open the file it hands you.

I’ll try to remember all of this the next time a video gets me excited. No promises, though.

Thanks for sticking with me through a post about a dog in a hat and a lot of exports.

See ya 👋

😏 Don’t miss these tips!

We don’t spam! Read more in our privacy policy

Related Posts

Leave a Comment

Your email address will not be published. Required fields are marked *