What "breakdown" means in production
A breakdown is the bridge between the screenplay (creative document) and the production (logistical undertaking). Every element the production will have to acquire, schedule, or coordinate gets tagged so it can flow downstream into the shooting schedule, the budget, the call sheet, and the shot list.
Industry convention is twelve categories, each with a colour: red (cast), yellow (stunts), pink (atmosphere), green (extras), blue (props), violet (wardrobe), brown (set dressing), orange (special effects), light blue (vehicles), magenta (animals), olive (music), tan (sound). A manual breakdown highlights every mention of each in every scene. An AI breakdown does the same — faster.
How AI breakdown actually works
Four steps. Split the script into scenes on the INT./EXT. regex. Process scenes in parallel batches of four against a fast LLM with a structured-output prompt asking for tagged JSON. Deduplicate and normalise across scenes ("MAYA" in scene 4 and scene 12 is the same character). Persist to a breakdown_elements table with FKs to the project and scene. Downstream queries — "how many scenes does Maya appear in" — become one-line lookups. The whole pass, on a feature, takes 30–90 seconds.
What AI breakdown does well
Three things, reliably. Speaking parts: character names are syntactically distinctive in a screenplay (uppercase, centred above dialogue). An LLM finds them all. Explicit props: "Maya picks up the revolver" is unambiguous. Every modern LLM tags it. Scene-level metadata: INT./EXT., day/night, location, page count, estimated screen time. All structured signals an LLM extracts cleanly.
What AI breakdown misses
Three things, also reliably. Implied props: "Maya pours herself a drink" implies a glass, a bottle, a counter. An LLM catches the glass maybe 70% of the time; the counter almost never. Background action: "the crowd cheers" is one stunt tag, but can hide twenty unwritten beats. Director intent: a scene says "the chase ends at the bridge". The director wants "a steadicam handheld 360° spin". That's not in the script. It has to come from the shotlist + the director's bible.
Pattern: AI breakdown gets you 80% there in 60 seconds. The remaining 20% is a careful human read.
What to check before you trust an AI breakdown
Five-minute audit. Character roster (does it match what you remember?). Vehicle list (catch the Uber in scene 14 and the chase truck in scene 31?). Stunt density (gates insurance and bond requirements). Wardrobe per character (look for under-counts). Special effects (rain, fog, fire, prosthetic blood — under-tagged because often implied).
What this looks like inside StoryboardCanvas
The /breakdown route ships a full-script AI pass that batches scenes 4 at a time against GPT-4o-mini, writes 12-category industry-coloured tags into the production database, and shows per-scene success/failure so a single transient model error doesn't fail the whole breakdown — you only pay for scenes that actually completed.
The tags flow downstream automatically. Characters seed the cast list. Props seed the props inventory. Vehicles seed the vehicles inventory. Wardrobe links into character costumes. Schedule reads the breakdown to suggest day order.