Edit your videos with AI: the good-take workflow

What ChatCut is, and what it can do 🎬

It's a cloud video editor with a real multi-track timeline, and the plugin lets an AI agent drive it for you. The editing doesn't happen inside the chat: your agent sends instructions to ChatCut, and the timeline lives in your browser at app.chatcut.io. That matters because you can always open the project and finish by hand.

The same plugin works in the ChatGPT desktop app, Claude Code, Cursor, Grok and Kimi, so you're not locked to one assistant.

Install the plugin 🔌

Three routes, all documented by ChatCut. Pick one:

1. ⚡ The one-liner (fastest) Open the ChatGPT desktop app and paste this into any task. It installs itself in about a minute.
   paste into the desktop app

   /goal Read chatcut.io/chatgpt to install the ChatCut plugin and set up a new task for me.

2. 🔎 From the plugin store Open the plugin store in the desktop app, search ChatCut, click Install.
3. 🛠 Manually Plugins in the sidebar → Create → Add marketplace → paste https://github.com/ChatCut-Inc/agent-plugin.git → open the Personal tab → find ChatCut → Install.
4. 🔑 Sign in when it asks It opens ChatCut's authorisation page in your browser. Sign in and allow. Never paste passwords or tokens into the chat itself.

What it costs 💳

Mark your good takes out loud 🎬

You will do the same line five times. The agent can see all five and has no idea which one you liked. So tell it while you're still rolling.

After a take you're happy with, say a marker out loud in your normal voice, then carry on. ChatCut's own blog describes beta users doing exactly this, saying "ChatCut, mark this as director's favorite," then adding to the prompt: "Never use the parts addressing 'ChatCut' in the final video, but use them as logging notes." That second half matters, so include it.

Cut the speech first, everything else second ✂️

This is the single most useful rule I found, and it's in ChatCut's own guidance: the speech timing set by your A-roll cut anchors everything downstream. If you ask for cuts, captions, graphics and music in one prompt, you cannot tell which decision went wrong when it comes back odd.

So go in passes. Their own copy-paste prompts, in order:

Use ChatCut to make a clean first cut of this talking-head video. Remove repeated takes, filler words, obvious restarts, and long empty pauses, but preserve the speaker's meaning and natural cadence. Keep the strongest opening line, add readable captions, and leave all media and edits on an editable timeline. Do not add generated footage or music yet.

Remove repeated takes, filler words, obvious restarts, and unnecessary pauses. Keep the meaning intact, preserve any pause that helps the point land, and make the delivery feel natural.

Add word-level captions and highlight the word currently being spoken. Keep the text readable on a phone, away from the speaker's face, and consistent throughout the video.

Add subtle sound effects, transitions, and punch-in zooms only where they connect ideas or emphasize an important line. Keep the existing cut unchanged and avoid an effect on every sentence.

One more worth keeping, for turning a long recording into shorts:

Find three self-contained highlights in this recording. Turn each into a separate 9:16 short. Open with the strongest complete line, preserve enough context for each idea to stand alone, keep each timeline editable, and add phone-readable captions. Do not invent quotes or cut a sentence in a way that changes its meaning.

Teach it your look 🎨

Here's where I have to correct something I believed myself: you cannot hand it links to your old videos and have it learn your style. That isn't a feature. What exists is better documented and easier:

1. 🖼 Build a Design Style Open Design Style → Create Design Style. It takes colours, fonts, logos, reference images and written rules. Screenshot the frames from your own videos you want it to copy and upload those. No scraping tools, no ffmpeg, just screenshots.
2. 🔤 Pick fonts from its catalogue Not from your machine. ChatCut warns that system fonts like Arial or Helvetica fall back unpredictably between computers. Custom fonts attach to your Design Style.
3. 🔊 Use the built-in sound library first There's a sound effects library, and their own guidance is to use it before generating anything, because generating costs credits and the library doesn't.
4. 💾 Save it as a Skill Once a workflow works, open a project, click the book icon under the message field, and choose Save this editing process as a Skill. It comes back under My Skills in your other projects.

What it does well, and what still needs you 👩‍💻

Where it actually breaks. Independent coverage is genuinely thin, so this is the one substantive outside critique that exists, and it matches my own experience: the agent "can still struggle to interpret more subjective stylistic choices, pacing nuances, or complex narrative intents," and auto-picked B-roll "can feel repetitive and generic." ChatCut itself concedes that transcription "misreads proper nouns, brand names, technical terms, and fast or accented speech," which matters if you say tool names on camera.

The failure worth bracing for, from an editor who hit it: "I asked an editing agent to 'make this punchier,' got a tight cut back, and then realized it had removed the one sentence that made the whole point make sense."

Motion graphics and little animations 🎨

This is the part most people miss. ChatCut writes motion graphics as code, which means the animated stat, title card or lower third it makes for you stays editable afterwards. You can ask for a change instead of starting again. Ask in one pass, after the cut is locked, so the graphic lands on timing that won't move.

Add a motion graphic to the timeline at [the moment I say the number], on top of the existing footage, and do not change the cut underneath it.

It's a stat card that counts up to [87%] with the label [of viewers never make it past 3 seconds]. Vertical 1080 by 1920, about [3] seconds, positioned in the [upper third] so it never covers my face.

Style: one accent colour [#D97757] on a [cream] card, everything else neutral, one clean sans-serif. The number counts up with an ease-out, the label fades in slightly after it, and the whole card leaves the way it came in. No bouncing, no spinning, nothing pulsing while it sits there.

Show me a preview frame before you render it, and keep the graphic editable so I can change the number later.

Your own little mascot animation 🦀

A small looping character is a cheap way to make a video feel like yours, and you don't need a design tool. Ask your agent to draw it in code as pixel art, so every version afterwards is a number change rather than a re-roll. You need Python and one free library (pip3 install Pillow), and your agent can set that up for you.

Make me a small animated pixel mascot in Python using Pillow. Draw every frame in code from square blocks. Do not generate images with AI.

Character: [a round orange cat with two dot eyes]. Action: [typing on a laptop, looking pleased with itself].

Rules: build it from chunky pixel blocks so it reads as 8-bit, and write one function that draws the character with arguments for the parts that move, then call it once per frame so the character never changes size or colour between frames. Every frame is transparent, with no background. Pick one ground line and keep the feet on it, so it never floats. Around 16 to 24 frames for a simple loop. Export one looping GIF with a transparent background, saved as [name].gif.

Then show me the first frame and tell me the frame count. I'll tell you what to change and you adjust the numbers and export again.

Expect the first working version in about ten minutes, then a couple of minutes per tweak. Drop the GIF over your footage in ChatCut, or in whatever you edit in.

The free tools that do the boring parts 🛠️

You don't need any of these to use ChatCut, but they're free, they run on your own machine, and an AI agent can drive all of them. This is the layer that costs you nothing and saves hours.

1. ✂️ auto-editor, the silence cut Cuts dead air and filler words from a recording, and can export a timeline straight to Premiere, Resolve or Final Cut instead of a finished video. Install with brew install auto-editor, then run auto-editor video.mp4 --margin 0.2s. Public domain, genuinely free.
2. 🎧 ffmpeg, for everything else Converting, cropping to 9:16, normalising loudness, burning in captions, exporting a transparent overlay. Nobody remembers the syntax, which is exactly why handing it to an agent works so well. Install with brew install ffmpeg-full.
3. 📝 Transcription on your own machine The full ffmpeg build can now transcribe audio to an SRT file by itself, and brew install whisper-cpp gives you a faster one built for Apple Silicon. Free, offline, and nothing gets uploaded.
4. 🪄 rembg, for cutouts Removes the background from an image in one command: pip install "rembg[cli]", then rembg i input.png output.png. Handy for thumbnails and overlays.

The ceiling, and what's past it 📈

ChatCut's motion graphics are genuinely good for captions, simple animations and basic charts, and you can edit the generated pieces afterwards. What it can't do is bind a chart to a real dataset or give you frame-exact control.

If you need that, the step up is building the graphic as code, where a chart reads from a real dataset and you control every frame. Two options worth knowing, and the licensing differs sharply. Remotion is not open source, it's source-available under a proprietary licence, but it's genuinely free for an individual, commercially, forever. That only changes at four or more people. HyperFrames is Apache-2.0 with no headcount threshold at all.

The learning curve is smaller than it was. Both now have coding-agent paths, so you can prompt your way to a first render rather than writing React from scratch. What you get back is still code, so every refinement means reading or re-prompting it. Reach for this only for the recurring branded asset, the chart or title treatment you'll rebuild every month. It's an overlay on your edit, not a replacement for it.

Does it replace Descript? 🤔

For the job most creators use Descript for, cutting a talking head down from the transcript, removing filler and dead air, and getting a tight first cut, yes, it genuinely does. And it adds things Descript doesn't have: an outside AI agent that can run the whole cut for you, motion graphics that stay editable as code, and generation of video, images and music inside the editor.

Rights and your footage 🔒