The problem
A small team of award-winning independent filmmakers had just finished a documentary and were turning a short film into a feature. Coordination ran on memory and goodwill, so mistakes got through. In one case, a character was left out of a day's setup.
Writing was its own struggle. They found it hard to get ideas onto the page, and then hard to see how the story was taking shape. When a detail changed late, two finished scenes had to be rewritten.
They couldn't afford a production assistant or a script editor. And they didn't want AI anywhere near the creative work.
Constraints
- No AI-generated creative content. It never writes scenes, characters, or story.
- Close to free to run, on hardware they own.
- No new tools to learn. The filmmakers work in Obsidian, Docs, Slack, and email.
- Nothing reaches cast or crew until a producer confirms it.
A team of production assistants
FilmAI is a set of AI agents, each with one job. None of them writes the film.
- Production managerDaily / weekly
- The orchestrator. It's built on a modified version of Karpathy's LLM Wiki pattern, restructured around projects. It sets up a folder and templates for every production, from future ideas to shoots in progress. Every day it checks each project for outstanding tasks and assigns them to the other agents.
- Script supervisorOn call
- A thought partner, not a writer. It takes in ideas in bite-sized pieces and files them against the right script. It understands scene structure, cross-references characters and locations, and tracks continuity and intent. When it spots a contradiction or a theme drifting, it questions the writer.
- Production assistantOn call
- Manages the shared Excalidraw board the team uses to see and coordinate the production, including storyboards and budget.
- Daily PADaily
- Builds the daily reports and shot list, and collects RSVPs from cast and crew.
- Location scoutOn call · in progress
- Tracks locations and helps find shots.
Under the agents sit code-only workers that do the mechanical work: sending email, parsing replies, posting to Slack, writing Docs, and building investor pitch decks in Slides. They call a model only when a step needs judgment.
OpenClaw runs the agents on a Mac mini M4 Pro (48 GB) in the filmmakers' house, using a local model. It launched on Qwen 3.5 in June and moved to Ornith 1.5 (35B A3B) in July. Claude Code runs only the script supervisor.
Decisions
A production assistant, not a content generator. The filmmakers were hostile to AI until it was framed as the assistant they couldn't afford, one that never touches the creative work. That framing was the adoption strategy.
Use a model only where judgment is needed. Code writes the emails. Code reads replies first, and a local model gets only the unclear ones. Claude handles the one job that needs real reasoning: talking through a script. That's why the whole system costs about $110 a month.
Story facts live in files, not in the model's memory. Characters and locations are cross-referenced files the system checks against. That's how it can tell the writer what a late change will affect.
Run it in parallel before trusting it. FilmAI ran alongside the old way of working, so we could see what it got right and what it missed before anyone depended on it.
How it fails, and what happens then
The daily schedule is the riskiest output, because it reaches people outside the core team. It goes through three gates:
-
Human
A producer reviews every schedule against the scenes being shot that day, before anything is sent.
-
Code
Code generates the emails, not a model, from the approved data, so nothing made-up reaches cast or crew.
-
Code → model → human
Replies are handled in tiers. Code handles clear replies and updates the record directly. Anything less clear goes to the local model, which returns YES or NO plus a comment. If the model can't tell either, the reply is flagged for the producer.
During the parallel run, a sound crew member was missing from a daily schedule. The producer caught it at review. The model wasn't the cause: that person had never been entered into the system. The fix was structural. Crew who are always needed are now flagged and selected before any schedule is generated, so the same gap can't recur silently.
Results
- In daily use by 4 producers and their crews across 4 productions.
- About $110 a month to run: $108 for four Claude seats and $1–2 in power. In heavy writing weeks they buy extra usage credits, or the writer moves up a plan.
- First-draft cycle went from about 50 hours to about 15, by the writer's own estimate.
- It catches what a script editor would. In a horror short, the first three scenes were deliberately oblique. The supervisor flagged the fourth as on-the-nose, and the writer rewrote it.
- Built in about two weeks to V1, and in production since June 2026.
What's next
- Finish the location scout.
- Keep iterating on what the producers report back.