Claude Code Edits Video From a Transcript: How the video-use Skill With 30 000 Stars Works

The open-source video-use project from the browser-use team turns Claude Code into an automatic video editor. Claude Code is a version of Claude that runs on your computer and can open files, launch programs and create new files on its own. The video skill (in agent terminology, a "skill" is a package of instructions and scripts for a single task) appeared on GitHub in spring 2026, is released under the MIT license and has collected more than 30 000 stars in six months.

You put the source footage in a folder, write what to remove and what to add, and a few minutes later a finished video sits next to it. The skill also works with OpenAI's Codex, but the installation instructions are written for Claude Code.

Why a text model can handle video

Claude does not open the whole video file. It works with a transcript. Picture a printout of an interview in which every word is marked with the time it was spoken. You cross out the "um"s, repetitions and botched phrases, and an assistant cuts the recording according to your marks. The pipeline works the same way:

• The recording is transcribed into text with the exact time of every word. Pauses, laughter, applause and speaker changes are marked.
• Claude reads the text and decides what to keep.
• If a spot is unclear (a long pause, two similar takes), it looks at a few frames from that segment.
• Its decisions are written as a list of splices: "from 0:12 to 0:47, then from 1:03 to 1:30."
• FFmpeg, a free program for working with video, assembles the video from that list.
• Claude checks the joins and, if it finds errors, redoes the edit, with up to three attempts.

Token economics: gigabytes become 12 kilobytes

The main engineering idea is compression. All takes are packed into a single file, takes_packed.md, of about 12 KB. That is a few thousand tokens, which fit easily into the model's context. For disputed spots, the skill builds a timeline_view image: a storyboard, the audio waveform and word captions. Decisions are stored in an EDL (edit decision list), and the project state lives in project.md, so you can pick the work up again in the next session.

browser-use applied the same trick to browser control: the agent gets not a screenshot of the page but a structured text description of it. Here, a transcript with timecodes takes the place of the DOM.

Cost depends on the length of the recording. Transcription is billed by the minute, and Claude spends more tokens on longer text. A one-hour project is noticeably more expensive than a ten-minute one, and on a Pro subscription it is easy to hit the limit. Claude Code requires a Pro subscription (from $20 a month) or Max, or an API key with pay-per-token billing. For users in Russia, the article suggests an option to pay in rubles via SpeShu.AI, which offers the Claude Opus 5.5 and Claude Sonnet 5 models.

What it can do, where it is weak and what goes to the cloud

The strong side of the editing is speech:

• removes filler words, slip-ups and long pauses;
• cuts between words, not in the middle of a sound;
• keeps the best take when there are repeats;
• applies a 30 ms audio fade at every splice to avoid clicks;
• does color correction, subtitles and title cards.

Claude programs animations rather than drawing them. There are three tools for this:

• Remotion: product videos, interfaces, charts (video is described as React components);
• Manim: formulas, diagrams, scientific explanations (scenes written in Python);
• HyperFrames from HeyGen: animated graphics described as a web page (HTML and CSS, rendered through headless Chrome).

The skill does not create cutaways with other footage: you need to put them in the folder in advance.

The weak spots follow from the architecture. A request like "place the shot right after he smiles" requires understanding the image, and Claude sees only text and individual frames. Transcription errors carry over into the edit, especially in noisy recordings, with strong accents and in Russian. The model does not replace a sound engineer: it removes noise only to a limited degree and does not mix music or effects. You need a terminal to run it.

The video itself does not leave your computer, and the splicing is done by local FFmpeg. What does go over the network is the audio track (if you choose cloud transcription), the transcript with your instructions and a few disputed frames. For recordings under NDA there is the freecut fork: it replaces cloud transcription with Whisper running on the user's machine. By default that is mlx-whisper on Apple Silicon and faster-whisper on other systems. freecut has about 250 stars against 30 000 for the original, and updates reach it with a delay.

Who benefits

It suits anything where the main element is a person speaking: interviews, podcasts, lectures, breakdowns, demo videos. For a product demo, screenshots and Remotion are enough. It is better to write prompts in English, because the skill's instructions are in English. A good request answers five questions: which files and in what order, what the result should be (duration, format, quality), what to cut, how to style it, and which inserts to add and when.

Watch the video yourself before publishing. Claude checks the technical side of the joins, but the meaning is still your responsibility: clipped words, misspelled names in subtitles, volume jumps, and removed pauses that were part of the delivery. Write your corrections with timestamps.

There is also a licensing trap. video-use itself is under MIT, and commercial use is allowed. But Remotion is free only for individuals and teams of up to three people.

The takeaway for engineers and product teams: the "transcript → LLM → EDL → FFmpeg" pipeline is reproducible. The prompt, the decision list and the animation code are all plain text, so you can keep them in git and use them as templates. Quality depends on transcription accuracy and on how well the model holds a long context. Claude does not replace specialized video neural networks: it solves a different problem, but does so quickly. If you regularly publish talking-head videos, the savings are measured in hours per week. If you edit once every six months, a familiar editor is still more convenient.