Title: Artie Spiegel — AI Showrunner

Inspiration

Every AI writing tool has the same shape: you type, it finishes your sentence. That's fine for an email. For a screenplay it's a problem, because the U.S. Copyright Office requires human authorship, the WGA agreement says AI output isn't literary material, and a writer who let a model draft their pages may not own what comes out.

So we asked a different question. What if the AI ran the process and never touched the page?

What it does

Artie Spiegel is an AI showrunner. He interrogates before you write, diagnoses after you write, and never writes anything himself.

The Greenlight fills twelve slots through conversation — protagonist, want, flaw, antagonism, thematic proposition, ending shape. Complete it and Artie moves from skeptical to committed, the title page formats itself, and you land on scene one.

The Scene Rig runs before each scene: seven slots, reset every time, clearing instantly on deterministic checks so nothing blocks the writer.

The workbench formats screenplay components as you type. Press Enter after a character cue and you're in dialogue. Press Enter on an empty cue and you fall through to action.

The Script Supervisor scores each scene against a 72-cell matrix — twelve narrative positions by six craft lenses.

The Director boards a scene from its action lines alone, then reports what the model filled in where your description ran out.

How we built it

Three agents, deliberately flat. Artie is never the parent of the others, because ADK delegation would pass scene prose down the tree and break the reading boundary.

Google Gemini on Cloud Run. ClickHouse for the matrix and the provenance ledger. Confluent between them. Supabase for transactional state. Built with IBM Bob.

Before Artie speaks he deliberates across four opposed axes — expansion against restriction, persistence against yielding. Both poles argue with a strength rating, synthesis decides what to raise and what to hold, and the whole trace is visible and stored.

Every keystroke is captured as a hash-chained record. Not the characters — deltas and a chain of hashes proving the text grew by typing.

Challenges we ran into

Making the firewall structural rather than instructed. An instruction can be jailbroken; an absent field cannot. Artie's payload has no place to put prose, and an assertion raises loudly if one appears.

A calibration constant we got wrong by reasoning. We set the provisional-input factor to 0.5. Against real data it inverted the priority ordering — a weight-5 attested gap ranked below a weight-3 one. It's 0.75 now, chosen by data rather than argument.

The image model never admits it's guessing. Six words of description produced an old man, a fedora on his lap, and a clock reading 3:45 — all confident, none requested. A fidelity check was impossible, so the feature became a diagnostic that reports what got filled in.

What we learned

The constraint is the product. A tool that refuses to write for you is only valuable to someone who needs to have written it themselves — and that turns out to be a real person with a real problem.

We also learned to concede what we couldn't solve. The matrix models one narrative tradition. Kishōtenketsu fails its conflict column. A survival film with almost no dialogue is systematically disadvantaged. We measure departure rather than judging it, and we say so.

What's next

Mobile reading. C2PA credentials on generated frames. Cross-scene continuity checking. And a marketplace — because a script with a verified authorship record is a script you can hand to a buyer without a provenance conversation.

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Title: Artie Spiegel — AI Showrunner

Built With

• clickhouse
• cloud-run
• confluent
• css
• docker
• fastapi
• fountain
• gemini
• google-adk
• google-cloud
• google-genai
• html
• ibm-bob
• javascript
• kafka
• nano-banana-pro
• postgresql
• pytest
• python
• screenplain
• supabase
• uvicorn