How to edit videos with Claude: what it really does and what it doesn't

You have a two-hour interview and need a three-minute video plus six vertical clips by Friday. You drag the file into Claude, type "edit this for me", and the answer is that it can't open it.

That is not a problem with your account. As of October 2026, Claude does not accept video files: it works with text, documents and images. According to Anthropic's documentation, it takes JPEG, PNG, GIF and WebP images — up to 20 per message on claude.ai — and reads only the first frame of an animated GIF.

That does not make it useless for editing. It means you have to give it the video in a form it can read, or give it tools that do the cutting. This guide is for editors, creators and agency teams who cut video every week and want to know which part of the job Claude can do today, without the hype.

The five real ways to use Claude for editing

What you give it	What Claude does	Where	What it's for
The timestamped transcript	Decides what to cut and in what order	Regular Claude (chat)	Selects, paper edits, clips
Individual frames	Looks at and describes images	Regular Claude (chat)	Picking thumbnails, checking framing, continuity
Your terminal and FFmpeg	Runs the cuts and exports	Claude Code	Cutting, captioning, reframing and batch export
A Remotion project	Writes the video as code	Claude Code + Remotion	Motion graphics, data videos, templates
An MCP connector	Talks to another app	Claude with connectors	Searching your footage archive or driving an editor

Most useful workflows combine two of these. Let's go through them.

1. Editing from the transcript

This is the most profitable approach and the one that needs the least setup. Cutting an interview, podcast or talk is mostly deciding which sentences stay. Claude is good at that, because it is text.

The process:

1. Transcribe with timestamps. Any tool that exports SRT or VTT will do. If the material is confidential, the Polimake desktop app transcribes it locally with Whisper, without uploading it anywhere.
2. Give Claude the SRT and the goal. Don't ask it to "summarize": ask for concrete editing decisions.
3. Take the list to your editor and build a first rough cut from it, or hand it to Claude Code to cut (point 3).

A request that works:

Attached is the SRT transcript of a 1 h 52 min interview with the founder
of a skincare brand. I need:

1. A 3-minute video for the website about how the brand started.
2. Six vertical clips of 30 to 50 seconds for Instagram, each with a
   single idea that makes sense without context.

For each piece return a table with: start, end, the opening line, the
closing line and why you chose it. Only use timestamps that appear in the
SRT. If a cut leaves a sentence unfinished, tell me.

What works well: finding lines that stand on their own, proposing a narrative order, spotting repetition and writing the supporting text for each clip.

What you need to check: that the timestamps exist exactly as in the SRT — the last line of the request is there for that — and that the cut sounds right. Claude reads what was said, not how it was said: a perfect line on paper may have a cough in the middle or an intonation that doesn't land.

2. Reviewing frames

What Claude can see is images. If you pull frames out of the video, it can judge them. With FFmpeg, one frame every five seconds:

ffmpeg -i interview.mp4 -vf fps=1/5 frames/%04d.jpg

Uses worth the effort:

• Picking a thumbnail from twenty candidates, using criteria you give it ("face visible, looking at camera, clean background"). Then you need to test which one works, and Claude doesn't decide that.
• Checking a vertical reframe: whether the person falls out of shot when cropped to 9:16. The same goes the other way, when you adapt a vertical video to horizontal.
• Spotting continuity problems between shots from two shoot days.

The limit is clear: it sees still images. It doesn't perceive movement, pacing or audio. Don't ask it whether a cut "breathes".

3. Actually cutting with Claude Code and FFmpeg

This is where Claude stops advising and starts doing. Claude Code is the version of Claude that works on your computer: it reads files, runs commands and checks the result. With FFmpeg installed, it can cut the clips from step 1, burn in captions, reframe to vertical, normalize audio and export everything in a batch.

You describe the task in plain language and it writes and runs the commands. It asks for permission before running anything, so you see every command before it touches a file.

It is the approach that saves the most time on repetitive work, and the one that needs the most setup. We walk through it step by step, with the commands it should end up running, in editing video with Claude Code and FFmpeg.

You need a paid Claude subscription or an Anthropic Console account, and you shouldn't be afraid of a terminal window.

4. Generating video with code: Remotion

Remotion is a library that builds videos as if they were web pages: every frame is code. It has an official Claude Code plugin, so you can ask for "a 15-second animation with these four figures from the report and the logo at the end" and Claude writes the project.

It fits:

• motion graphics and data-driven videos;
• templates where only the text or figures change, a hundred versions at once;
• brand intros, lower thirds and end cards.

It doesn't fit editing recorded footage. That's what point 3 is for. And before using it on client work, check Remotion's license: it isn't free for every company.

5. Connecting it to other apps over MCP

MCP is the standard Claude uses to connect to other applications. For video there are two uses that make sense.

Driving the editor. There are community MCP servers for DaVinci Resolve that let Claude create timelines, add clips or start renders. They are not official Blackmagic products, each needs a different Resolve version and all of them require the app to be open on the same machine. Useful for automating repeated tasks; unreliable for cutting a piece from start to finish.

Finding the footage. The bottleneck in many edits isn't cutting: it's finding the shot, especially B-roll from old shoots. If your archive is in Polimake, you can connect it to Claude and ask, from the conversation, "find shots of hands applying the cream from this year's shoots". Polimake describes and transcribes what you upload, also searches inside videos and returns the link to the original. Claude doesn't edit those videos: it helps you decide which ones to use and finds them for you, even a shot for every line of the script.

What Claude still doesn't do well

• Judging pace. It decides on text and still images, not on how the piece feels to watch.
• Audio. It hears nothing. Filler words, noise or music only show up if they are in the transcript, and it doesn't replace an audio mix.
• Color and finishing. It can apply a LUT with FFmpeg if you give it one, but it doesn't replace a color grade.
• Made-up timestamps. Without a timestamped transcript it can suggest cuts at moments that don't exist. Always require it to use only timestamps from the file.
• Confidential material. A client's transcripts and frames leave your computer when you use the chat. Check the contract first.

The whole workflow in one app: the Polimake desktop app

If you've read this far, you've seen the workflow needs several pieces: something to transcribe, something to find the shots, an editor and a way for Claude to reach your files. The Polimake desktop app puts them together on your computer:

• Local Whisper transcription, with no per-minute cost and without the video leaving your machine. That's the SRT from point 1.
• Search across your folders by what is said, what is shown and what appears on screen, without uploading anything.
• An open-source video editor, with no subscription or watermark, to cut what Claude suggests. We compare it with others in free video editor with no watermark.
• MCP and a command line, so Claude can query your library and Claude Code can work with the files.

Download the Polimake desktop app for free: none of this costs anything. Try point 1 with a video you already have. For now there is only a Windows version.

Where to start

If you have never tried it, start with point 1 on a video you've already cut yourself: transcribe it, ask for the selects and compare them with yours. In half an hour you'll know whether its judgment saves you work or creates more.

If you like the result and you cut the same thing every week, the next step is Claude Code with FFmpeg. To see which other parts of the edit you can hand off, even outside Claude, read editing videos with AI. And if what eats most of your time is finding the footage, start by organizing your raw footage: no assistant edits well with shots nobody can find.

Frequently asked questions

Can Claude watch or analyze a video?

Not directly: Claude doesn't accept video files, only text, documents, and images. To have it work with a video, you give it the transcript with timestamps or frames extracted as images. With Claude Code, it can also run FFmpeg to cut and export the files.

Which AI assistant can analyze video directly?

Google's Gemini accepts video as input and can describe what happens in it. Even so, for deciding cuts precisely, working from a timestamped transcript is still more reliable, because it lets you check each timestamp against the original file.

What's the difference between an SRT file and a VTT file?

Both are text subtitle files with timestamps, and either works for handing a transcript to Claude. SRT is the format most compatible with editors and players; VTT (WebVTT) is the web browser standard, starts with the line WEBVTT, and separates milliseconds with a period instead of a comma.

How can I get a timestamped transcript of a video for free?

With Whisper, OpenAI's open-source transcription model, which exports SRT and VTT and can run on your own computer. The Polimake desktop app includes it: it's free for Windows and transcribes locally, without uploading the video to any server.