From script to shopping brief

The problem

On a television production, the wardrobe department runs on the script breakdown: every scene, every character, every costume, tracked by hand across shooting schedules that never run in story order. Each script revision restarts the work. Continuity mistakes — a jacket that changes between scenes shot on the same day — are expensive to reshoot. And the paperwork extends past set: procurement lists, quick-change planning, stunt-double duplicates, customs manifests for international shoots.

What we built

A chat-driven copilot for the costume team. Designers upload screenplays, shooting schedules, and casting sheets; a deterministic parser — built on published academic screenplay-parsing research — structures scenes and characters in under two seconds, then AI enrichment extracts wardrobe mentions, continuity signals, and action flags in the background.

From that knowledge base, the agent produces working deliverables, not chat answers:

• Per-scene wardrobe breakdowns, rendered as reviewable artifacts.
• Shopping briefs that specify fabric, color, and quantity, informed by each character's arc — a character written as wealthy-and-declining gets different materials than one on the rise.
• Continuity alerts ranked by severity, and quick-change and stunt flags.
• Concept art grounded in the actual actor's measurement sheet.
• Customs manifests generated from photos of packed wardrobe cases.
• Museum-sourced research moodboards with full attribution.
• Revision diffs showing exactly what changed for a character between drafts.

The harness pattern

The domain knowledge lives in roughly 27 swappable tool modules around a general orchestrator. Every model call routes through a single registry and gateway with provider fallback, and every image generation is governed by a versioned, model-independent prompt contract — so the system survives model deprecations by design. It is a model-agnostic harness, currently deployed on Gemini.

The same chassis was later instantiated for a completely different industry: recruiting.

Verifiable facts

• Structural script parse in under two seconds, verified by integration tests against real screenplay PDFs.
• Around 27 domain tools, each a typed, schema-validated contract.
• Before shipping image generation, we ran an internal review covering actor-likeness rights and cost controls — the kind of diligence generative media needs.