VIOLET Studio

The first was shape. VIOLET's thesis — research a brand, then hold every generation to it — is an editorial idea, but it lived in a generator. Brand consistency is something you judge across a sequence, and there was no sequence: you generated, downloaded, and assembled the work somewhere else, where the brand profile no longer existed.

The second was keys. VIOLET's front end reached the model providers directly, so the API keys compiled into the JavaScript bundle — which is why it can never be hosted anywhere public. Studio's first milestone was a hard architectural rule, and it holds today: no key is ever sent to or read from the browser. Generation goes through a server-side proxy, and the brand research doesn't use a key at all.

VIOLET's original argument was a system that holds an opinion about its own output. Studio finishes it twice over: the same brand profile that shaped the prompt is used to grade the result, and the timeline the result lands on is exported by a local pass that places every clip at its timestamp and in-point — so the file is the cut. Press play, or step through it; every box is a real module and every number in it was measured.

This is the thesis, running. The profile on the left was produced by the app's own research route against the NBA, for the spec spot above, and the composition on the right is a port of lib/brand.ts. The dial is not a fade — it is a ladder. Each threshold folds one more layer of the brand into the prompt, in the order the source uses.

Three details in there are the difference between a demo and a tool. The palette is emitted as colour language plus hex — royal blue (#1D428A) — because a diffusion model treats a bare hex code as close to noise while the colour name actually steers it. The brand's don't-list is routed to the negative prompt with its imperatives stripped: writing “avoid crowd shots” into a positive prompt makes crowd shots more likely, and leaving the word “avoid” in a negative prompt negates the negation. And a rival's mark is excluded at every active level, not just the high ones — it used to open at 0.5, which is how a swoosh once ended up in a demo.

The first thing a fresh install shows is a 35-second spot the studio made about the studio: six clips and two stills generated in-app, a music bed and a voiceover generated on the audio lanes, twelve overlays and the VIOLET mark on the GFX lane — cut on the same timeline, rendered by the same export, and pinned in the project menu with a reset. It opens once, and never hijacks an install that already has work in it.

Work proceeded in order, and a milestone was not done until its acceptance checks passed — the same SPEC discipline as QUARRY and DETOUR. The criteria are the interesting part, because they are all observable: “a 17s edit shows a ~17s ruler”; “pick a brand → generate 3 stills → visibly consistent styling”; “a stranger can npm i && npm run dev with no keys and see a working timeline”.

Every fix is logged with what broke, why, what changed, and the measurement that verified it — 84 entries at the last count. One milestone was reworded mid-flight and signed off: brand research moved from “an OpenAI key server-side” to the keyless CLI, and the doc records that it changed and why. A late discovery got its own line too: routing generation through a provider's prepaid credits turned out to be impossible as designed, because its API bills a separate wallet from its subscription — blocked, written down, then built with the constraint understood rather than discovered twice.

// the same profile that shaped the prompt grades the result
auditMedia(image, brand) → { paletteScore, styleScore, verdict, issues, fix }

// runs automatically, but only when a brand actually shaped the generation,
// so unbranded work costs nothing extra. Failures are swallowed on purpose:
// an audit is commentary, and must never mark a good generation as broken.

164 TypeScript files32,823 lines15 API routes 72 commits7 working days