One Prompt, 18 Visual Styles: How I Proved My Short-Video Identities Are Actually Different
I run a few faceless short-video channels (TikTok, Reels, Shorts), and they all had the same problem: they looked like everyone else's. Same template, same captions, same stock motion.
So I asked an AI agent to explore the opposite: how many genuinely different visual identities can you build for vertical shorts, entirely in code - and can you prove they're different?
The result is an open-source repo with 18 identities, each proven by 2-3 real shorts rendered by code (45 in total), and a single prompt that turns any of them into your own studio.
StudioVisualIdentity - un prompt, 18 styles, votre chaîne 18 identités visuelles pour vos vidéos courtes, prouvées par 45 vraies vidéos. Un prompt suffit pour donner son style à votre chaîne. Open source. h4md1.fr
TL;DR
- 18 visual identities for 1080×1920 shorts - weather map, 1900 newspaper, chalkboard, risograph zine, 16-bit pixel flyover, cut paper…
- Each is a design system an agent can read: palette roles, type pair, motion, transitions, safe areas, recipes per content family.
- 45 proof shorts, all rendered frame by frame by code (D3, GSAP, three.js, p5, Manim, Taichi, numpy…), voiced, mixed and checked by a QA script.
- One prompt gives an agent everything it needs: pick a style, scaffold a separate studio, render episode 1.
- The surprise: my automated "are these styles distinct?" metric said 41/45. A blind LLM judge said 72/72. The metric was the one that was wrong.
The problem with "make me a style"
Ask an agent for "a cool visual style for my channel" and you get one style, described in adjectives, with nothing to check it against.
What I wanted was a menu with evidence: several identities side by side, each one documented in the same format and backed by real renders, so the choice is made by looking, not by vibes.
So every identity in the repo ships the same things:
identity.md- always the same 16 sections, in the same order (promise, palette roles, typography, motion language, camera, transitions, short structure, safe areas, libraries, suitable / unsuitable content, risks…)tokens.jsonwith a normalised core validated against a JSON schema (colours, @fontsource slugs, durations in seconds, safe areas in px)recipes/<family>.md- how to open, structure and close a short of that family in that style- 2-3 proof shorts with their source,
narration.json, stills and ashort.jsonthat records the exact command to reproduce them
All of it is indexed in one machine-readable file. This is what an agent actually reads:
{
"id" : "carte-meteo",
"promiseEn" : "A clear weather map where isobars breathe, wind runs in dark trails, temperature pills pop, and a weather front sweeps the screen diagonally to move on.",
"familiesProven" : [ "evenements", "geographie" ],
"coreTokens" : {
"color" : { "bg" : "#CFE6F5", "surface" : "#FFFFFF", "text" : "#10263D" },
"safeAreaPx" : { "top" : 220.0, "bottom" : 320.0, "right" : 120.0, "left" : 60.0 }
},
"curation" : {
"audience" : [ "adults", "teens" ],
"signatureToPreserve" : [
"diagonal cold-front (blue triangles) / warm-front sweep as the only transition",
"text only on white cartouches"
],
"confusableWith" : [ "swiss-data" ]
},
"renderCostSecPer100Frames" : { "min" : 18.1, "max" : 18.7, "measuredUnderLoad" : true }
}
signatureToPreserve and confusableWith are the two fields that matter most. They tell an agent what it must not touch when it adapts a style, and which neighbour it would drift into if it did.
One prompt, one studio
The repo is the exploration; you don't fork it to run a channel. A script scaffolds a separate project from one identity - your draft identity, scene templates for the content families you need, the render toolkit and an operating manual for the agent.
The README boils it down to one prompt:
Read AGENTS.md from https://github.com/killerwolf/StudioVisualIdentity. Pick the style that fits a [topic] channel for [audience] in [language]. Create our studio with out/tools/bootstrap_studio.py, adapt the visual identity to our channel, then make our first episode about [episode topic] and show me the result.
Under the hood, the agent ends up running something like this. Here's a real run from today - a geography quiz channel for teens, built on the weather-map style, borrowing a template from the arcade style for a family the weather map never proved:
python3 out/tools/bootstrap_studio.py ../geo-quiz --channel-id geo-quiz --name "Geo Quiz" \
--from carte-meteo --blend pixel-arcade --families geographie,engagement \
--language en --audience teens
studio créé : …/boot2/geo-quiz
template geographie <- carte-meteo__geographie__le-mistral-pourquoi-il-souffle-si-fort
template engagement <- pixel-arcade__engagement__quelle-bille-gagne-la-course
(template emprunté à pixel-arcade : référence de STRUCTURE (rythme, mécanique), pas de signature - à restyler entièrement avec vos tokens)
suite : lire AGENTS.md dans le studio (npm i, .venv, puis out/tools/new_episode.py)
Yes, the tooling talks French - the project started for French channels.
Note the warning on the borrowed template: a blend keeps one identity's signature and only borrows structure from the other. Mixing two signatures is exactly how you end up generic again.
The catch: my distinctness metric was wrong
Here's the part that surprised me. To check that the 18 styles really are different, the repo has a measured test: a colour histogram in Lab space (Bhattacharyya distance) plus an edge histogram (χ²), 8 frames per short, and for each short, "is its nearest neighbour a short of the same identity?" It scores 41/45.
The four misses:
- liquid-chrome · Newton's cradle - atelier-papier (cut paper)
- liquid-chrome · four MPM paintings - neon-lab
- liquid-chrome · which colour fills the glass - cosmos-doc
- survol-pixel · how an atoll is born - aquarelle-conte (watercolour)
Taken at face value, that says liquid-chrome isn't a coherent style at all.
So I ran a second test the way a viewer would: a fresh agent with no context got 72 anonymised frames - 2 shorts × 2 frames for each of the 18 identities, random file names, the key kept out of its reach - and one instruction: make 18 groups of 4 by style, not by subject. It made 18 pure groups out of 18, 72/72 frames.
Its own notes put the finger on why the metric failed: the liquid-chrome group "joins two quite different scenes (test tubes vs Newton's cradle), linked by the shared glossy iridescent 3D rendering and black ground", and the dark identities (neon lab, cosmos, topo relief, pixel arcade) "are neighbours in palette but separated cleanly by type and rendering technique".
That's the lesson. A colour-and-edges metric sees palettes. A style is mostly geometry and type: the chrome reflections, the pixel grid, the perspective of a 3D flyover. Two shorts of the same style on different subjects can have very different palettes (a blue ocean and a green mountain, for the pixel flyover), and two different styles can share a black background.
I kept the metric - it's cheap and catches regressions - but the blind judge is now the evidence that counts.
What it doesn't do
It's a style menu and a production toolkit, not a channel in a box. Topic writing, fact-checking, review, publishing and scheduling are still yours (my own channels run those in a separate private template piloted from Notion; this repo only feeds it the style).
And the agent can't hear: it measures loudness and checks timings, but someone still has to watch and listen to every episode.
Rendering is CPU-bound and uneven: from under 5 s per 100 frames (cut paper, single-line drawing) to around 3 minutes for the three.js space documentary, and over 5 minutes for one Manim chalkboard short - measured with several renders sharing a 4-core machine.
Wrapping up
Everything is open source - the 18 design systems, the 45 shorts and their code, the render/voice/QA tools, and a Claude Code skill (fabrique-identites) that runs the whole "scout → adversarial challenge → studio → review" loop to add new identities.
Landing page and demo: https://h4md1.fr/StudioVisualIdentity/
Repo: https://github.com/killerwolf/StudioVisualIdentity
If you build a studio from it, I'd love to see which style you picked.
References
out/report/recognition.jsonandout/report/comparison.md§4 in the repo - both tests, raw resultsout/identities/core-tokens.schema.json- the token schema every identity followsAGENTS.md- the operating manual the prompt points the agent to
Comments
No comments yet. Start the discussion.