Make Codex Watch AND Transcribe Client Videos
Codex can already watch a video and describe what's happening. What it can't do out of the box is transcribe what's being said — ask for that, and it'll point you to the OpenAI API, another paid subscription. Here's a free, open-source fix.
The idea
A client sends you the task as a video instead of text — a Zoom call, a screencast, whatever. Before, that meant a lot of manual work: pausing, screenshotting, typing it all out by hand.
Codex can now watch that video for you and describe what's happening. But it doesn't analyze the voice — ask it for help with that, and it'll suggest the OpenAI API instead. Another paid subscription just to get a transcript.
Faster-Whisper is a free, open-source tool that transcribes audio locally — no API key, no extra subscription, and it's extremely fast. The prompt below plugs it into Codex, so one skill both analyzes what's happening in the video and transcribes it with timestamps.
The prompt
Paste this into Codex to build the skill. It installs Faster-Whisper if it's missing, runs it on your video's audio, and combines the visual analysis and the timestamped transcript into one file.
Create a new skill called "video-to-brief" that does two things whenever I give it a video file:
1. Analyze the video itself - describe what's happening on screen, step by step, in the order it occurs (Codex can already do this natively, so just call that built-in video analysis).
2. Transcribe the audio track using faster-whisper:
- Check if faster-whisper is installed; if not, install it (pip install faster-whisper).
- Run it on the audio from the same video file, using the "small" model for a good speed/accuracy balance.
- Output the transcript with timestamps (segment-level is fine), formatted as [MM:SS] text per line.
3. Combine both outputs into a single markdown file called task-brief.md, structured as:
- A short summary of what the video shows and what's being asked
- A "Visual timeline" section (from step 1, key moments with rough timestamps)
- A "Transcript" section (the full timestamped transcript from step 2)
Save task-brief.md in the same folder as the input video. Make this a reusable skill I can run on any video file going forward - not a one-off script.Need help adapting AI tools to your workflow?
I help developers and companies put tools like Codex and Claude Code to work in real, everyday tasks — not demos. Message me and let's talk about yours.