# An agentic film studio: brief to storyboard to shots to finished cut Entry E15. Sector: Prospicience internal studio. Status: internal. Delivered: 2026-09. Source: https://ai.prospicience.in/work/agentic-film-studio Prospicience's agentic film studio takes a written brief to a finished cut: storyboard, character sheets and keyframes, shots, music and voice, then assembly. A person writes the brief and approves the finals, a director agent plans, a runner agent executes, each shot goes to the model best suited to it across four providers, and a cost gate approves spend before anything runs. Speed: First finished cut in under two hours. ## The challenge Making a short film with AI meant picking a model for every shot by hand, watching over each generation, and losing finished clips when provider download links expired hours later. ## Why it mattered Hand-picking models eats the director's time, and a wrong pick shows up in exactly the shots that matter most. Expired links meant paying for the same generation twice, and without a spending check, exploration could use up the budget before a single final shot existed. ## What we built - Built a pipeline from brief to storyboard to stills to shots to music and voice to an assembled cut. - Routes each shot to the model that suits it: one for dialogue realism, another for physics and motion, a cheaper one for drafts, an identity locked one for stills. - Downloads every generation the moment it finishes, before the provider link can expire. - Adds a cost gate per run, with a draft tier and a final tier so exploration stays cheap. ## The result - A repeatable studio that turns a brief into an assembled film with generated music and sound. - Every paid output is kept on disk, logged with the model, prompt and cost behind it. - Drafts run cheap, and only approved shots are made at final quality. - Characters hold across shots in reference conditioned sequences, as the second cut on this page shows. ## Built for trust Nothing is generated until the run passes its cost gate, and every output is saved the instant it finishes. Queued work is never silently regenerated, so every charge on the card maps to a shot someone asked for. ## AI at work AI generates the storyboard, character stills, shots, music and voices for a short film, with a director agent planning each run and a runner agent carrying it out. A person writes the brief, approves the spend and chooses which drafts go to final quality. What it produces: Storyboards from a written brief, Character sheets and keyframes, Video shots, each from the model that suits it, Music, voice and sound effects, Assembled cuts. - Each shot goes to the model that does it best, so the director stops hand picking models shot by shot. - Drafts run cheap and only approved shots are made at final quality, so the budget goes into the film rather than the experiments. - Character sheets, locations and settings are reused across shots, and every output is logged with its model and cost, so each production starts from what worked last time. - The first finished cut came together in under two hours. How it works day to day: A person writes the brief. The director agent turns it into a storyboard, the runner agent sends each shot to the right model once the cost is cleared, and every file is saved the moment it finishes. The person reviews the drafts, picks the finals, and the studio assembles the cut. ## Figures from the delivered system - providers routed: 4 - pipeline stages: 6, brief to cut - quality tiers: draft and final - generated files retained: all, downloaded on completion ## The technology - Describe the film, get a storyboard - Each shot goes to the model best suited to it - Music, voice and sound effects generated alongside - Everything assembled into a finished cut - Nothing is spent before you approve the run AI models used: Seedance, Veo, MiniMax, Kling, Nano Banana, Seedream, Flux, GPT Image, ElevenLabs, Suno. ### Technical notes - Director agent pinned to a high reasoning model for planning, runner agent on a faster model for execution - Per shot model selection across four providers by capability: dialogue realism, physics, cheap drafts, identity locked stills - Structured storyboard schema so shots, references and audio cues stay machine readable through the pipeline - Immediate artefact download on completion, since provider URLs are short lived - ffmpeg assembly with normalisation across differing model outputs - Per run cost estimation gate before any provider call ## Questions ### Why route across several video models instead of picking one? Because each has a different strength. One holds a face across shots, another handles physical motion, another is cheap enough for drafts. Routing per shot gives every shot the model that does it best. ### How do you control the cost of AI video generation? A cost gate before the run and two quality tiers. Exploration happens at draft quality where a mistake costs very little, and only approved shots are generated at final quality, so the budget goes into the film rather than the experiments. ### What happens to generated files? They are downloaded the instant they finish and logged with the model, prompt and cost. Provider links expire, often within hours, so nothing already paid for is ever lost. Capabilities: ai-media-generation, ai-agents-and-automation.