Open-Write is a precision instrument for designing long-form stories — screenplays, novels, television. Like CAD for an architect or Blender for an animator, it holds the entire structure in view, checks it against itself across hundreds of pages, and renders your draft. The design, the judgment, the soul of the thing stay yours.
It is not a button you press to get a novel. It is a tool a craftsperson drives. An architect using CAD is still the architect; the instrument supplies precision and throughput, never taste. What you make of it is up to you.
The work is readable now · the open-source repository releases June 1, 2026
Long-form coherence was never a model problem — it's a scaffolding problem. Open-Write organizes standard, general-purpose AI models into specialized roles and supports them with structured state, so the work holds together across the full arc of a season, a novel, or a screenplay. You move scene by scene through the same stations a real production would.
Three templates ship with the public release — film screenplay, novel, and episodic television — each self-contained, with its own bible structure, mode definitions, and export tools. A state server holds the things a writer loses track of across three hundred pages: who knows what and when, which plants have paid off, what the audience believes, where each scene sits on the clock. And because same-model critics share blind spots, evaluation runs across more than one model and takes the union of what they flag — the way you'd never have a draft proofread only by its own author.
Yes — you can hand it a premise and it will draft the whole structure. A Thousand Silences was made that way, from a single prompt, with no hand-editing at all. But be clear about what comes back: a buildable draft — the developed structure, not the finished, consecrated building. It is the cathedral roughed in, not the cathedral made worth standing in.
Closing that gap — the judgment about what the story should feel like, the taste that separates a draft from a thing worth reading — is still the architect's work. The instrument supplies throughput. It does not supply that.
Three completed projects are released alongside the system as evidence of what it can do in skilled hands. A Thousand Silences is readable in full now; excerpts of the others are posted ahead of the June 1 release.
A novel of the Haitian Revolution, following a French translator who slowly recognizes that the precision of his work is what makes him complicit. Drafted from a single prompt — a setting, a historical source, a literary inspiration — with iterative pipeline tuning and zero editorial intervention.
Readable in full, free, today.
Read all 30 chapters →A hard-science metaphysical drama about a grief-frozen physicist contacted by the quantum-realm successors of all advanced life. Three completed forms — feature screenplay, novel, and an expanded screenplay covering the full Velai civilization in a parallel storyline.
RECOMMEND — from the system's own AI adversarial reader.
Read an excerpt →A four-season historical drama about John Brown and the Bleeding Kansas period. Spec scripts for all four seasons; three drafted in full, the fourth in development. A tragedy in the older sense, about the cost of conviction and the question of unpaid debt.
In development for studio submission.
Read the Season 2 cold open →If you have a problem with AI in creative work, you're not wrong to ask hard questions. Here are mine, answered the way I'd want them answered — without dodging the ones that don't have a clean answer.
Is this just slop? Did a machine make it?
It was made by AI, yes — and the way to settle whether that makes it slop is to read it, not to argue about the tool. The instrument doesn't supply taste; the person driving it does. An architect working in CAD is still the architect, and no one calls a building drafted in CAD "not really architecture." Judge the building, not the pencil. So read the novel cold and decide for yourself.
Isn't this trained on writers' work, taken without consent?
The models underneath are third-party, general-purpose, and pretrained — I didn't build or train them, and I won't pretend the question of how they were trained is settled, because it isn't. It's a real, unresolved fight, and I don't claim to have solved it.
What I can answer for is what I built on top of them, and the choices that were actually mine: releasing the whole system openly instead of selling it, putting the demonstration works into the public domain, and being honest about who this costs rather than pretending it costs no one.
Won't this put writers out of work?
It will pressure specific roles — developmental work, spec work, junior story roles especially — and I won't pretend otherwise. But the capability is arriving whether or not I build my version of it. What I could control was whether it lands in everyone's hands or only in the hands of whoever can pay the most for it. I chose the first. The people in the exposed roles deserve honesty and a head start, not reassurance that nothing is changing.
Then what's left for the writer?
Everything the instrument can't supply: the judgment about what a story should feel like, the taste that separates a buildable draft from a finished one, the conviction underneath the whole thing. The tool rewards the people who bring those. That's why I hold a few components back from the public release — so the floor doesn't fall to where pure button-pushing clears it. The throughput is free. The judgment is still yours, and still rare.
I'm Nicholas Detweiler — a writer and systems builder. A month ago I finished reading a novel I had not written. I'd asked a system of AI agents I built to produce one from three pieces of direction: the Haitian Revolution as setting, Laurent Dubois's history as a source, Hamlet as an inspiration rather than a template. What came back changed what I thought was coming for this work.
The common knowledge was that AI loses the thread over long arcs — fine for a paragraph, hopeless across a book. It turned out the problem wasn't the model. It was the scaffolding. With the right division of labor and the right structure underneath them, ordinary models can carry an entire novel. I had no say in whether tools like this would exist. I did have a say in how mine would be built and shared — toward broad access or narrow capture. I open-sourced it. I kept a small set of components private, not to gatekeep, but because releasing them would drop the floor far enough to flood the field with work that passes a glance and means nothing — which helps no one and hurts the working artists this is meant to serve.
I write quickly and I think in public. If you want the longer argument — what this capability is, why it's here sooner than the industry believes, and what I think we owe each other as it arrives — that's what the essays are for.
The system is free; anyone can learn to drive it. But if you'd rather hand the drafting to someone who built the instrument — and who brings the components that aren't in the public release — that's what I do.
Have a manuscript you want developed into a screenplay, a story you want carried to a full draft, a season you need built out? I take it through the full pipeline and bring the judgment and the private components the public tool doesn't include.
Fixed scope · fixed deliverable · by projectA workshop for writers and small studios who want to use the instrument themselves — bible to finished draft, the parts that don't come through in documentation. Built for people, not engineers.
Cohort workshop · in developmentIf you're a production company or press thinking seriously about AI-assisted development as part of a real pipeline — not a gimmick — I consult on building it well, with quality control at every station.
Engagement · by conversationOpen-Write is open-source and the demonstration works are in the public domain — no paywall, now or later. I built it on two months away from regular work. If it was worth something to you, you can put something in the jar. Entirely optional, always.
Support Open-WriteThe full public pipeline — three templates, the Python toolchain, the state server, and the methodology docs — released under Apache 2.0.
Why the danger in this transition is structural rather than individual, and why the right unit of analysis is the incentive structure, not the people inside it.
The Structural Solution, followed by the closing coda, Designing for the Creature We Actually Are.