Project Studio

The problem

Film tycoon games simulate production beautifully and then turn the film itself into a score. The one thing they can't do without a language model is make the film real.

So I wrote the promise down before writing code: in Game Dev Tycoon you can't play the game you make; here you can read the film you made.

What I built (so far)

Stage 1 is a vertical slice to test that one idea before building anything else:

• Pick a logline (or type your own), then genre, tone and ending.
• The engine writes the film as 8–14 readable scenes, each with a picture. You can cut a scene or rewrite one in your own words.
• A writers' room gives notes on a board, then critics review the film by quoting specific scenes.
• Top-down pixel art, with keyboard and controller support.

The studio map, casting, budgets and progression are designed but not built. I'm only building the next stage once the current one feels right.

How it works

• All scenes in one call. The whole scene list is generated together, so the story holds together for free instead of being stitched from separate requests.
• Numbers never come from the model. Box office, budget and consistency are plain JavaScript, so they are deterministic and testable in Node.
• Written to port. Game rules are pure functions that never touch the DOM or the canvas; everything on screen is drawn on a canvas. So a later Unity or Xbox build means rewriting the renderer, not the game.
• Planned: fully offline. The design splits AI work into generation (a small local LLM with grammar-constrained JSON) and judgement (a small classifier exported to ONNX, which runs in the browser and in Unity). Not built yet.

What went wrong / what I learned

My first version told the player the right answer. Before release, a single "script editor" said things like "sold as a mystery, it plays more like a horror". That turned a creative choice into an optimisation problem: players stop asking "what do I want this film to be?" and start asking "what scores best?".

The fix: the room now has several people who disagree. Two advisors contradicting each other isn't a hint, it's a decision. Scores only appear after release, when they are a consequence rather than a guide.

Instant box office felt fake. Knowing the audience and sales on release day killed the tension, so I wrote down why release needs time to pass, and what that costs the scoring, before changing any code.

Built with an AI coding assistant as pair programmer; the design, the architecture and the testing are mine.

Stack

• JavaScript
• HTML Canvas
• LLM APIs
• Gamepad API
• Systems design

Links