Claude vs ChatGPT Video Editing: Same Clip, Same Prompt

We filmed our founder, Allen, talking to the camera for 79 seconds and gave the raw clip to two AI agents with one written prompt: Claude Opus 5.5 in Claude Code, and GPT-6 Astra in Codex, ChatGPT's coding agent. The edit orders were spoken inside the clip itself. "Write my name behind me." "When I snap my fingers, make a clone of me." Each agent had to hear the order and fire the effect at that moment.

Both finished, and both landed every order without anyone touching the result. Where they split was in how they read one vague word, how much computing power they spent on cutting Allen out of the background, and how hard they checked their own work.

Below are the setup, the results order by order, the full prompt, and a step-by-step guide to doing it yourself.

Can Claude or ChatGPT actually edit a video?

Yes, with one condition: the AI is not clicking around in an editing app. Claude Code and Codex are agents that work inside a folder on your computer and run programs there. For a video, that means they transcribe the audio, find the moments that matter, cut you out of the background, draw graphics and render a new file. You describe the result in plain English, and they write and run any code themselves.

So the quality depends on how clear your instructions are and which free tools the agent picks. The prompt handles the first. The test shows the second.

How we set up the test

1. The clip: 79 seconds, filmed in portrait at 4K and 30 frames per second, 2,382 frames. Allen improvised the orders, and the only edit before the test was removing his false starts.
2. The orders, spoken in this order, each followed later by "remove it" or "back to normal":
   • Write my name behind me, and make it huge.
   • When I snap my fingers, make a clone of me, and make him copy me.
   • Make everything black and white except me.
   • Make my background look like the Matrix, with code rain falling behind me.
   • Put a news overlay on the video, like I am a speaker on a news channel.
   • When I open my hand, make words fly out of it.
3. The prompt: identical for both, word for word. Only the output folder names differed.
4. The rules: each agent worked alone, one after the other, never at the same time. Each had to deliver a file with exactly the same length and frame count as the raw clip, so the two could play in sync, plus an EDIT-NOTES.md file explaining each order, its time and the tools used.
5. Afterwards: nobody touched either edit. We stacked the two files and published them side by side for people to vote on.

The results, order by order

What both got right

• Both transcribed the clip with faster-whisper large-v3, a free local version of OpenAI's Whisper, using word timestamps.
• Both cut Allen out with Robust Video Matting, a free cut-out model, so his name and the code rain could sit behind him.
• Both drew the graphics with HyperFrames and made the final file with FFmpeg.
• Both delivered exactly 2,382 frames at 30 fps, audio at -14 LUFS, with no cuts and no speed change.
• Both made their own music bed in code instead of downloading a track.

Where they split

The clone is the clearest difference. The order was "make him copy me". Claude read "copy" as a delay, so the copying is visible. Astra built a mirror that moves at the same time as Allen. Both are fair readings of the same sentence, which is the most useful lesson of the test: one vague word can send two capable AIs in different directions.

The cut-out is the second. Claude ran the larger model on every frame at full 4K on the graphics card, an NVIDIA RTX 5070. Astra ran a lighter model at 1080x1920 on the processor. We did not measure edge quality, so the honest summary is that Claude spent more computing power on it.

The third is the self-check. Claude ran an automatic test on every frame that no text touched Allen's face or hands, looked at every 10th frame on contact sheets and checked each trigger closely, and listed ten problems it fixed along the way. Astra looked at every frame on 40 contact sheets and larger stills at each effect, and noted openly that it measured the audio but did not listen to it.

Who won

We let viewers decide. Of 128 comments we read the day after the side-by-side went up, about half asked for the prompt, and almost all the rest voted for Claude or Opus. Our own read: both edits are usable, and Claude's gesture timing and delayed clone are the clearest differences you can see.

The free prompt: edit a video from your spoken orders

This is the public version of the prompt both agents got. Our file paths and machine settings are swapped for placeholders. The job, the color rules, the text and sound rules and the self-check are the same.

CopyEdit my video. In the video I say the edits out loud. Do each one at the exact moment I say it.

INPUT:  [the path to your raw video]
OUTPUT: [where to save the finished video]
STYLE:  [optional: a reel whose look you want, or delete this line]

THE JOB
- My spoken lines are edit orders. Transcribe the video with word timestamps (local Whisper works well), find every order, and do it the moment I say it. Sync it to what I do on camera, like a snap, a point or an open hand. Check the frames, not only the words.
- An effect stays until I say "remove it" or "back to normal".
- My first lines are the intro and my last line is the call to action. Make them look great too.

MAKE IT EXCITING
- If I gave you a STYLE reel, watch it first and use its look: colours, captions and full-screen graphic cards.
- Then go further: faster, bolder, more surprises. Something new should happen every second or two.
- No italic words.

COLOUR: VERY LIGHT, MAKE ME LOOK GOOD
- Grade very lightly. Keep it natural.
- Make my skin look healthy and warm. Take any coloured light from screens or LED strips off my face.
- Do not push contrast or saturation. Do not crush the shadows.
- No skin smoothing, beauty filter or vignette.

TEXT AND SOUND
- Captions: 1 to 3 words at a time, bold sans, never on my face or hands. Big enough to read on a phone.
- Clean my voice. Add a music bed and short sound effects, all under my voice.

HARD RULES
- Same length and timing as my video. No cuts, no speed change. Frame 1 of yours = frame 1 of mine.
- Same resolution and frame rate as my video. H.264, audio at -14 LUFS.
- Use free tools that run on my computer.

BEFORE YOU FINISH
- Check every frame of your output. Look for: an effect early or late, broken cut-out edges, text on my face or hands, typos, blank frames, sound out of sync. Fix them, render again and check again, until you find nothing.
- Write EDIT-NOTES.md next to the output: each order, its time, what you did and the tools you used.

Work end to end without asking me questions.

Prefer a one-click copy? The same prompt is on our free prompt page.

Why each part is there:

• THE JOB makes the agent transcribe with word timestamps and then check the frames, so an effect lands on the snap or the open hand itself.
• COLOUR keeps the grade light and rules out skin smoothing and beauty filters. Removing screen light from your face is the one correction it asks for.
• HARD RULES keep your timing intact, which is what lets you compare two edits frame for frame or drop the result into a longer video.
• BEFORE YOU FINISH makes the agent inspect its own output. Both agents used it in our test: each rendered, looked at the frames, fixed what it found and rendered again.
• The last line tells the agent not to stop and ask questions, so it works through the whole job in one run.

Before you start

The free tools the agents reached for, so you know what they install:

• faster-whisper by SYSTRAN, MIT license. A rebuild of OpenAI's Whisper that its makers say runs up to 4 times faster for the same accuracy, with a timestamp for every word.
• Robust Video Matting, built at ByteDance by Shanchuan Lin and colleagues. It cuts a person out of ordinary video with no green screen. It is GPL-3.0 licensed, which matters if you build it into a product.
• HyperFrames by HeyGen, Apache 2.0, which turns HTML and animation into video.
• FFmpeg, the free video tool almost every pipeline uses for the final encode.

Step 1: Film a clip the AI can edit well

1. Put the phone or camera on a tripod and do not move it. Clones and background swaps fall apart when the frame moves.
2. Stand or sit in front of a plain wall, wearing a top in a different color from it. The cleaner the edge between you and the wall, the cleaner the cut-out.
3. Leave empty space on both sides and above your head. The clone needs room next to you, and your name needs room behind you.
4. Keep your hands in frame at chest height for any order that depends on a snap, a point or an open hand.
5. Say each order clearly, then pause one to two seconds while looking at the lens. That pause is where the effect plays.
6. Say "remove it" or "back to normal" when you want an effect gone.
7. Film in short blocks and keep your best try of each. Then trim the false starts, because the prompt keeps your timing frame for frame and will not cut anything.

Step 2: Set up the folder and open the agent

1. Make a new folder and put the trimmed clip in it. If you want a look copied, add a video whose style you like.
2. In the Claude desktop app, click the Code tab, choose Local in the environment selector, and pick the folder with the project folder control. In the terminal, cd into the folder and type claude instead.
3. Type /model and pick Opus 5.5 if it is not already selected.
4. In the ChatGPT desktop app, choose Codex, start a new chat, open the folder, and pick GPT-6 Astra as the model if your plan has it.
5. Leave the permission mode on Manual for your first run. Claude then asks before it runs commands, including installs, and you approve each one.

Step 3: Paste the prompt and let it work

1. Copy the prompt and replace the INPUT and OUTPUT lines with your file names, for example INPUT: raw.mp4 and OUTPUT: edited.mp4. Delete the STYLE line if you have no style video.
2. Paste it and press Enter.
3. Approve the installs and commands it asks for. Expect transcription first, then the cut-out of every frame, then the graphics, the sound and the final render.
4. Let it run through its checks. It will render, look at frames, fix problems and render again. Neither agent recorded how long its run took, so plan for a long session and keep the computer awake.
5. When it finishes, you get the new video and EDIT-NOTES.md next to it.

Step 4: Review and ask for fixes

1. Watch the whole video once at full screen, then again with EDIT-NOTES.md open next to it.
2. Check that each effect starts on your word or gesture and ends on "remove it" or "back to normal".
3. Read every caption that shows a name or a product. Whisper wrote "GBT" for GPT in our test, and both agents corrected it. It also heard "code rain" as "range", which Claude caught.
4. Ask for fixes in exact terms: "The clone appears late. Start it on the frame my fingers snap, and make it copy me half a second behind." Vague notes get vague fixes.
5. When you are happy, ask the agent to delete its working files, or remove the scratch folder yourself.

Run your own head-to-head

To compare agents yourself, copy the clip into two folders, open Claude Code in one and Codex in the other, and paste the identical prompt in each. Run them one after the other, so they do not fight over the same processor and graphics card. Because the prompt locks the length and frame count, you can then ask either agent to stack the two results into one video, one above the other, and watch both handle each order at the same moment.

Variations for other jobs

Swap or add lines in the prompt. For example:

CopyFor a calmer look: delete the MAKE IT EXCITING section and add "Keep graphics simple and on brand. One effect at a time."

For your brand: STYLE: use my brand colors [your hex codes] and the font [your font].

For captions only: "Do not add effects. Only clean my voice, grade lightly and add captions."

For a precise clone: "When I snap, make a mirrored clone that copies my moves half a second later."

Limits and gotchas

• Whisper places some words early. Claude found that a word after a pause was placed about 0.4 seconds before Allen said it, and checked the timing against the audio waveform. If an effect fires early, ask the agent to check that moment against the sound.
• "Copy me" can mean two things, as the clone showed. Say exactly what you mean.
• Colored light on your face. Monitors and LED strips leave purple or blue light on skin. The prompt asks the agent to remove it, and a lamp on your face while you film leaves less to fix.
• Both agents kept their working files after finishing. Clear them once you are done.
• Licenses. Robust Video Matting is GPL-3.0. That is fine for editing your own videos, and worth checking if you plan to build it into software you sell.

Other ways to edit video with AI

• Descript lets you edit a video by editing its transcript, which is faster for cutting a talk down. It has a free plan, with paid plans listed from $16 to $65 per person a month. Pick it if you mostly need cuts and captions.
• A regular editor plus an agent. Let Claude Code build only the graphics and captions, then assemble the final cut in your usual editor. It takes longer, and you keep full control.

If you want this kind of video work done for your brand instead, that is what our creative production team does.

Who this is not for

• If you need a 10-second trim today, a normal editing app is faster.
• If your footage is handheld, shot outside or full of people, the cut-out effects will struggle. Talking-head clips on a tripod are where this works best.
• If you will not watch the result closely before posting, skip it. The self-check catches a lot, and you still need to look at every effect yourself.

Frequently asked questions

Can Claude edit videos?

Yes, through Claude Code. It works inside a folder on your computer and runs free tools there: Whisper to transcribe, a cut-out model to separate you from the background, HyperFrames for graphics and FFmpeg to render the final file. You describe the edit in plain English.

Which is better at video editing, Claude or ChatGPT?

In our test both delivered every spoken order on time from the same prompt. Claude Opus 5.5 ran a full 4K cut-out, timed the words to the exact frame the speaker's hand opened and made a clone that copied him a few frames late. Almost every vote in the comments we read went to Claude.

Do I need to know how to code to edit video with Claude Code or Codex?

No. You film yourself, put the clip in a folder, open Claude Code or Codex in that folder and paste the prompt. The agent writes and runs any code itself. You review the result.

Is AI video editing with Claude Code free?

The editing tools it uses are free and open source. Claude Code itself needs Claude Pro ($20 a month, or $17 a month billed yearly) or a higher plan. Codex is included in every ChatGPT plan, but GPT-6 Astra, the model in our test, rolled out to paid plans from Plus up.

Why do the effects fire at the right moment?

The prompt tells the agent to transcribe your video with word timestamps, so it knows the exact second you said each order. It then checks the frames for the gesture, like a snap or an open hand, so the effect lands on the action and not only on the word.

What kind of video works best?

A talking-head clip filmed on a tripod against a plain wall, with your hands in frame and a one to two second pause after each spoken order. Moving cameras and busy backgrounds make the cut-out effects break.

Sources

• Anthropic: Claude Opus 5.5 announcement
• 9to5Mac: GPT-6 Astra comes to ChatGPT and Codex
• faster-whisper on GitHub
• Robust Video Matting on GitHub
• HyperFrames on GitHub
• Anthropic: Claude Code in the desktop app
• Claude pricing
• ChatGPT and Codex pricing
• Descript pricing