Creative Software · AI SaaS

Sweatbox

Script-to-storyboard animation studio with canon consistency built in.

A browser production studio that kills canon drift — a context engine injects locked lore into every script, storyboard, and voice generation, so characters stay consistent shot to shot.

Private BetaActive development

Sweatbox

AI Video

By the numbers

Prisma models

Backend services

AI providers

105

Commits

The challenge

Creators lose hours juggling disconnected AI generators, and the result is shifting character designs, broken lore, and inconsistent visual styles. Nothing keeps the universe coherent as it moves from script to storyboard to voice.

The solution

A production-graph architecture — Franchise → Production → Episode → Scene → Shot — backed by a custom AI Context Engine that injects locked canon into every model prompt, so consistency is enforced by the system rather than the creator's memory.

sweatbox · context-engine.ts

// Sweatbox · canon injected into every prompt
const ctx = canon.for(shot)        // locked wardrobe + lore
image.generate({
  prompt: preamble(ctx) + shot.action,
})                                 // consistent across shots

Illustrative: the context engine injects locked canon into every model prompt, so characters stay consistent shot to shot.

Technical deep-dive

Architecture

Next.js 14 App Router, tRPC v11, Prisma ORM with a 17-model Supabase PostgreSQL database, and Supabase Storage in a monorepo layout.

Multi-provider AI

An agnostic AI service layer supporting Anthropic Claude, OpenAI, ElevenLabs voice, and HeyGen video, so each stage uses the best tool.

Writers' Room

Eight specialized AI personas — Director, Screenwriter, Lore Keeper, Dialogue Editor, and more — give canon-grounded feedback on drafts.

Character consistency

Wardrobe references plus prompt preambles enforce visual consistency across image generators, preventing drift between shots.

Product decisions

Model the whole production as a graph so canon has a real structure to attach to.
A context engine injects locked canon automatically — consistency is not manual.
Provider-agnostic so image, voice, and video each use the strongest model.
Living bibles as first-class data, not documents that fall out of date.

Built with

Next.js 14tRPC v11PrismaSupabase / PostgreSQLAnthropic ClaudeOpenAIElevenLabsHeyGen

Honest statusPrivate beta, active development.

Related guides

AI engineering · 5 min

Keeping AI on-brand: voice profiles, canon, and shared context

Generic AI output reads generic. The way to make it sound like you — or stay true to a fictional universe — is to make consistency a system feature, not something a person has to remember every prompt.

Read the guide →

Want something like this?

Tell us what you’re building, the code you have, or the problem you need solved.

Start a project