ChatGPT Can Now Edit Videos For You

Drag a raw talking-head clip into Codex, paste one prompt, get back a captioned, tightened, animated cut ready to post. Here's the setup, the prompt that actually works, what happens under the hood, and the honest first-run timeline.

Codex is the part of ChatGPT that writes and runs code on your computer. Remotion is a tool that makes videos out of code. Put them together and you get an editor that never opens a timeline: it transcribes your clip, finds every silence, cuts them, lays captions over the words as they're spoken, animates the moments that matter, and renders a finished MP4 into a folder on your Mac or PC.

I dragged a two-and-a-half-minute talking-head clip in, pasted the prompt below, and got back a captioned, tightened, animated edit. Here's the part a thirty-second video can't show: what's happening while you wait, what your computer needs, and how to word the prompt so you don't get a robotic edit or a wall of questions.

How it works, in one paragraph

Codex runs on your machine, so your video never leaves it. The Remotion plugin gives Codex a set of skills: how to start a Remotion project, how to caption, how to render. Codex writes a small program that describes your edit (which seconds to keep, where the captions go, when to zoom), then Remotion renders that program into a video using its own bundled copies of ffmpeg and a headless Chrome. Transcription runs locally with Whisper.

What it can and can't do

In the ChatGPT desktop app, click the switcher in the top left and choose Codex (the third option after Chat and Work). Sign in with your ChatGPT account if it asks. Codex opens with a project view instead of a chat.

Click Plugins, search Remotion, and click Install plugin. Or open the Remotion plugin page at www.remotion.dev/docs/ai/codex-plugin and click the install link there. There's nothing to configure. The plugin adds Remotion's skills to Codex: creating projects, captions, rendering, previewing in Studio, and its docs.

Remotion brings its own ffmpeg and headless Chrome, so you don't install those. Two things it can't bring: Node.js (version 18 or newer, from nodejs.org) and, on a Mac, the Xcode Command Line Tools, which Whisper needs to build. Open Terminal and run `node -v`; if you get a version number you're set. If you're not sure about Xcode, tell Codex "check that node and the Xcode command line tools are installed and install what's missing" and approve the prompts. This is a one-time thing.

In Codex, create a new project pointed at an empty folder (Desktop/edits is fine). Drag your video into the Codex window. Type `$remotion` in the prompt and accept the plugin when it suggests it. Now paste the prompt below.

HEADS UP: Codex asks permission before it runs commands on your machine. On the first edit expect a handful of "allow this?" prompts: cloning Whisper, building it, downloading a model, installing packages. Say yes to those. Say no to anything that isn't about this folder.

The first run, honestly

The video says thirty seconds. That's how long the prompt takes to type. The first edit on a fresh machine is twenty to forty-five minutes, mostly downloads and one render. Every edit after that is about ten, mostly render. Here's where the time goes so you don't think it's broken.

What Codex is doing while you wait: it converts your audio to a format Whisper reads, transcribes it with word-level timestamps, runs a silence detector across the track and builds a list of the seconds to keep. Then it writes the edit as code: one segment per kept chunk, captions grouped three or four words at a time with the current word highlighted, and a zoom plus headline on each key moment. Then it renders. If it's good, it shows you the transcript with the cuts marked before rendering, which is what the prompt below forces it to do.

The prompt that works first time

The one-liner works, but Codex comes back with four questions first: which platform, how long a silence, which transcription, what "animate" means. This version answers them and adds the approval gate. Swap the file name and brackets.

Edit [talking_head.mp4] for [Instagram Reels].

1) Transcribe it with word-level timestamps using local Whisper. If it isn't installed, download whisper.cpp and the medium.en model without asking.

2) Cut every silence longer than 0.6 seconds and every standalone "um" or "uh", leaving 120 ms of room tone on each side of every cut so breaths stay natural. No other audio processing.

3) Captions: 3 to 4 words per page, never a one-word page, bold white sans-serif with a black outline, current word highlighted in [#F5A623], block centered at 70% of the height, max 900 px wide.

4) Pick the 5 strongest moments (a number, a claim, a punchline), at least 10 s apart. On each: an 8% zoom easing in over 0.4 s and back out, centered on the speaker's face, plus a one-line headline under 24 characters that springs in at 22% height for about 2.5 seconds.

5) Output 1080x1920 at 30 fps even if the source isn't, H.264, AAC audio.

Before rendering, show me the transcript with every cut marked and the 5 headlines as a plain table I can edit. After I approve, render to outputs/[name]_edited.mp4, tell me the final length, and show me a still at each headline.

Why it's worded like that. 120 ms of room tone is what "keep it natural" actually means; without a number you get a machine-gun edit. Never a one-word page stops orphan captions. Centered on the speaker's face matters because a phone-framed head sits high and a center zoom drifts it. A still at each headline is the QA step: Codex can't watch the video, only read it.

The approval gate

Before it renders, Codex hands you something like this. Real words and times come from your clip; this is the shape.

# TRANSCRIPT talking_head.mp4 (160.4 s)
00:00.0 Okay so [cut "um"] I want to talk about
00:03.1 something nobody tells you. [cut 0.9s]
00:04.9 I spent forty thousand dollars in my
00:07.4 first year and made back eleven. [cut 1.4s]
...
# CUTS: 41 silences (28.6 s) + 9 fillers (3.1 s)
# 160.4 s -> about 128.7 s
# 5 KEY MOMENTS
1 00:04.9 "$40K in. $11K back."
2 00:31.2 "Nobody says this part"
3 01:02.0 "26 months. One lesson."
4 01:41.6 "The 3-second rule"
5 02:19.0 "Do it before you're ready"
Reply approve, or edit any line.

This is where you fix the transcript (your brand name, a number it heard wrong), swap a headline, or un-cut a pause you wanted. Say "change headline 3 to 'This took me 2 years' and render" and it does exactly that. Nothing renders until you say so, so a bad transcript never becomes a bad video.

You are the QA. Codex reads frames and numbers; it does not watch the result. Look at the stills it sends, then watch the MP4 once at full speed before posting. On a two-minute clip, expect to hand-fix two to five cuts that landed a hair early or late. Tell it the timestamps and it re-renders.

Five more edits

Same project, same chat. Each one is a follow-up, not a fresh setup.

The 30-second cut:

Make a 30-second version with only the best 3 moments, same caption and headline style. Include the sentence before and after each moment so nothing starts mid-thought. Hard cuts between them. Render to outputs/[name]_30s.mp4.

The 16:9 version:

Make a 1920x1080 version of the full edit. Pillarbox the vertical video at full height with a blurred, darkened copy of itself filling the sides (no crop, no upscale). Captions at 82% height, max 1200 px wide. Headlines at 14%. Render to outputs/[name]_16x9.mp4.

Jump-cut zooms:

Alternate the framing at every cut between 100% and 106% zoom so the jump cuts read as intentional. Keep everything else the same and re-render.

Progress bar and lower third:

Add a thin [#F5A623] progress bar along the bottom edge, and a lower third with [NAME] and [@HANDLE] that slides in at 2 seconds and out at 6. Re-render.

Batch ten clips:

Apply this exact edit style to every .mp4 in source/. Same rules, 5 moments each. One combined approval table, then render all to outputs/ and list the final lengths.

Music: drop a track in the folder and say how loud ("-18 dB under the voice, fade out over the last 2 seconds"). Codex can't source licensed music; for Reels, add the sound inside Instagram.

Limits, licensing, and the mistakes that stall people

Licensing: Remotion is free for individuals and small teams (companies up to three people). Bigger companies need Remotion's company license; check remotion.dev/license before this ends up in an agency workflow. Fonts from Google Fonts are fine for commercial use.