MUSE

AI Creative Director. From brief to production package in minutes

Comment

Every filmmaker starts with a vision. But turning that vision into a production-ready package — script, storyboards, style bible, shot list — takes days or weeks of pre-production work. Solo creators and small teams are especially bottlenecked here. We asked: what if an AI could do the creative direction work in minutes, not weeks?

I built MUSE because we live this problem every day. Bastion Harbor Studios has shipped 250+ generative AI projects across short films, branded content, and interactive installations. We know the gap between "great idea" and "production-ready package" — MUSE closes it.

MUSE is an AI Creative Director that takes a creative brief and produces a complete production package:

  1. Script — Structured screenplay with scenes, dialogue, camera notes, and emotional beats
  2. Style Bible — Comprehensive visual direction: color palette (hex codes), lighting style, camera choices, film stock references, and a Global Lock string for generation consistency
  3. Storyboard — Detailed frame-by-frame breakdown with composition notes, camera movements, and AI-optimized image prompts
  4. Creative Review — Multi-round AI critique that evaluates the package for coherence, quality, and creative merit
  5. Storyboard Images — AI-generated storyboard frames for every shot, styled to match the style bible

The entire pipeline runs end-to-end from a single text brief. No manual intervention required.

How I built it

  • Frontend: Next.js 15 + TypeScript — dark cinematic UI with real-time pipeline visualization
  • Backend: Next.js API routes calling Gemini directly via @google/genai SDK
  • Text Generation: Gemini 2.5 Flash for script, style bible, storyboard, and creative review (structured JSON output)
  • Image Generation: Gemini 3.1 Flash Image Preview for storyboard frame generation
  • Pipeline: 5-stage sequential pipeline with structured JSON contracts between stages. Each stage receives full context from all prior stages. Style bible includes a "Global Lock" string appended to all image prompts for visual consistency.

Challenges I ran into

  • JSON reliability across 5 stages required careful prompt engineering + response_mime_type: "application/json"
  • Style consistency across storyboard frames solved via the "Global Lock" technique
  • Serverless timeouts handled by streaming stage completions via SSE for real-time progress

What I learned

  • Gemini's structured JSON output is remarkably reliable for multi-stage pipelines
  • Multi-round AI critique significantly improves output quality — the agent catches issues humans miss
  • Separating text and image models allows optimizing each domain independently

What's next for MUSE

  • Voice-driven creative direction via Gemini Live API
  • Video generation integration (Kling 3.0, Veo 3.1)
  • Spatial Consistency Engine for environmental coherence across shots
  • On-chain agent discovery via ERC-8004

Built With

  • gemini-2.5-flash
  • gemini-3.1-flash-image-preview
  • google-genai
  • next.js
  • tailwind-css
  • typescript
  • vercel
Share this project:

Updates