Claude Code dissected someone else's TikTok frame by frame and built its own for $16

It's a familiar situation: you have a video in your head but can't put it into words. The brief ends up vague, and so does the result. The author of the Telegram channel Technocafe took a different approach: showing the agent a finished reference video and asking it to break down how it was made.

The reference was an October 28 TikTok from @vibecodingiseasy about building videos with code. It has 133 thousand views and almost 5 thousand saves, suggesting the format clearly works.

Claude doesn't watch video; it reads frames

An important caveat: Claude doesn't perceive video the way a human does. It has access to frames and text. Simply pasting a link and asking what's in it won't work. The sequence is to download the video first, extract the subtitles, and only then request an analysis. That gives the model two reference points: images and words tied to timestamps.

The analysis identified 46 cuts, a shot change roughly every three seconds, and a dialogue in which one character's questions voice the viewer's objections. This turns a "good-looking video" into a structure that can be reproduced.

The biggest mistake: one frame per second

The most common temptation is to extract one frame per second and assemble a contact sheet. That's enough for static imagery, but not for movement. A card in the reference video flies in over just four frames, so its motion simply doesn't show up on that sheet. The model learns that the card appeared, but not how.

The prompt therefore needs to explicitly demand a frame-by-frame examination of every transition: what moves, where it comes from, how many frames it takes, and which spoken word it coincides with. For animation built with code, those parameters are almost a complete specification: direction, duration, and synchronization with the voiceover.

The analysis then becomes three lists: what to use, what to change, and what to leave alone. The author stresses that the aim is to copy the recipe, not the video. The original creator's footage stays with its creator.

Workflow and cost

The order matters because mistakes are expensive. First comes the voiceover, then one still frame for each scene, and rendering only after revisions. Rendering this video took more than three hours, so catching a flawed scene in a still image is much cheaper than spotting it in the finished file.

Claude Code running Opus 5.5 handled the script, 3D scenes, and editing. The author estimates the cost at around $16 in equivalent API usage and 1.6% of the weekly limit. No video models or image generation were used: every frame was drawn with code in their own js-film engine. The voices were created in ElevenLabs; roughly 2600 characters cost about $0.30. The finished video features Claude and Codex arguing.

What to apply to your own work

If you make promos, demos, or explainer videos, the entire workflow transfers: download a reference, extract the subtitles, demand a frame-by-frame motion analysis, record the three lists, and only then write code. The same technique works for interface animation: instead of saying something "slides in smoothly," you give the agent specific frame counts and timing tied to spoken words.

There are limitations, too. Costs and limit usage are the author's own estimates, not independent measurements, and depend on the plan. The js-film engine is custom-built; the post offers no ready-to-use equivalent. A three-hour render makes iterations expensive, so still frames are essential. Copy the structure, not the assets: someone else's footage, voice, and music are off-limits. Finally, the post funnels readers into a paid club, so it's best to test the details on your own project.