One photo and one idea in, a 6 to 15 second paper-collage animation out, in the look of Vox explainer documentaries. Then every change is one more prompt: another person, another place, another object, the same animation.
Talk to them in their language. <dir> = video-agent/vox/<slug>/ in their project (slug: 2 to 4 words of the idea, kebab-case).
video-agent/vox/<slug>/
├── photo.jpg their photo, untouched
├── layers/ every layer of the collage (PNG with alpha, plates in JPG/PNG)
├── raw/ Codex renders before the chroma cut, and the job logs
├── shots.json the layers asked to Codex (name, kind, ar, subject)
├── animation.html the animation: window.setT(t), opens in any browser and plays
├── capture.mjs frames -> MP4
└── <slug>.mp4
1. Tools (once)
codex --version && codex login status
ffmpeg -version | head -1
node --version
- Codex missing:
npm i -g @openai/codex(Node 20+) orbrew install codex. Not logged in: they runcodex loginthemselves ("Sign in with ChatGPT"). Never ask for a key or a password in the chat. - ffmpeg missing:
brew install ffmpeg(macOS),sudo apt install ffmpeg(Linux),winget install ffmpeg(Windows). - Then in
<dir>:npm init -y >/dev/null && npm i puppeteer@24(headless Chrome for the capture, downloaded once).
2. The plan (one message, then their yes)
From the photo and the idea, plan the collage like a page of a magazine, then the moves:
- The hero: their photo, always the REAL photo (never redrawn by an AI: faces drift). Cut out (step 3) and printed as halftone by code in the page.
- A background plate: one full-frame 9:16 paper scene that sets the place and the time (a sky, a city, a map, a desk, a sheet of aged paper).
- 3 to 6 cut-outs: the objects of the story (a monument, newspapers, banknotes, a car, a building, clouds, birds). Each one is a separate layer so it can move on its own.
- The moves (6 to 15 s): the camera pushes in slowly the whole time, layers enter one by one every 0.5 to 1.5 s (slide in, rise from below, drop from above, a torn-paper wipe), something keeps living (clouds drifting, a bird flapping with two cut-outs, notes falling), one title on a torn paper strip if the idea has one (a place and a year, a name). All on a 12 fps stop-motion clock.
Show the layers as a numbered list (name, plate or cut-out, what it shows) and the beat timeline in 4 to 6 lines. Wait for a yes: each layer is one Codex image on their plan.
3. Their photo: cut out, never redrawn
cp "<their photo>" <dir>/photo.jpg
uvx --from "rembg[cpu,cli]" rembg i <dir>/photo.jpg <dir>/layers/hero.png
uvx comes with uv (curl -LsSf https://astral.sh/uv/install.sh | sh, or pip install uv). The first run downloads the model (about a minute). Look at the PNG: a clean silhouette, no chunk of background left. Then:
- Cut-out bad or no uv: keep the photo as a rectangular print instead, it is very Vox too:
layers/hero.jpgplaced as a print with a white border and torn edges (seereferences/animation.md, "The photo as a print"). - A person who never had a photo (Napoleon, a painting): an engraving or painting from the public domain (Wikimedia Commons), cut out the same way.
- Only photos of people who agreed, or public-domain images. Say it once if they send someone else's photo.
The halftone, the white scissor border and the shadow are added by the page (references/animation.md), so the real photo stays the real photo.
4. The layers with Codex
Write <dir>/shots.json and build every prompt with references/look.md: the subject, plus the green-screen sentence for a cut-out, plus the style sentence. Never the word "Vox" in a prompt, never a logo, a wordmark or any readable text in an image (newspapers and notes get illegible type texture only).
Run the jobs with the command of references/look.md, 3 at a time in the background, always ending with < /dev/null. A job takes 1 to 3 minutes. Then for each cut-out, the chroma cut:
ffmpeg -v error -y -i <dir>/raw/<name>.png -vf "colorkey=0x00FF00:0.30:0.10,despill=type=green:mix=0.6:expand=0,format=rgba" <dir>/layers/<name>.png
Check each layer by eye: put it on red to see any green fringe (ffmpeg -v error -y -f lavfi -i color=red:s=1024x1536 -i <layer> -filter_complex "[1:v]scale=1024:-1[a];[0:v][a]overlay=0:0" -frames:v 1 <dir>/raw/check-<name>.png), and no text or letters in it. A bad one: fix the subject and run that one again.
5. The animation page
Write <dir>/animation.html from the template of references/animation.md (the whole page is there, tested) and adapt it to the plan: the layers, their sizes and places, the timeline in setT. The rules that make it look like a Vox documentary are in the same file. Keep setT(t) a pure function of t: the capture calls it frame by frame.
Open it in their browser to preview (it plays on its own): open <dir>/animation.html (macOS), xdg-open (Linux), start (Windows). Fix what looks wrong before the capture.
6. The MP4
Write <dir>/capture.mjs from references/animation.md, then:
cd <dir> && node capture.mjs animation.html <slug>.mp4
About 4 s of work per second of video. Look at 3 frames (start, middle, end) before showing it: ffmpeg -v error -y -ss <t> -i <slug>.mp4 -frames:v 1 -vf scale=540:-1 raw/frame-<t>.jpg. Hand over the MP4 path. Music or a voice go on top in their editor, or with /short-editing.
7. Change it with one more prompt
Same <dir>, same page, only what changes:
- Another person ("do it with Napoleon"): a new hero (step 3), same layers, same timeline: the page swaps
layers/hero.pngand is captured again. - Another object or place: one new Codex layer (step 4), placed in the page, the rest stays.
- Another move or timing: only
setT. Another format: see "Formats" inreferences/animation.md.
Never regenerate layers that did not change.
Honest limits
- Codex's image tool decides the details: a layer can come with letters on it or a green fringe. Always look.
- The cut-out of a busy photo can miss hair or fingers: the print version is the safe fallback.
- Each layer counts against their ChatGPT plan's Codex limits.
- The look is inspired by Vox's documentaries: never their name, logo or wordmark on screen.