Nobody in a good scene says what they mean.
Screenwriting is the most constrained kind of writing there is. You get behaviour and dialogue and nothing else — no interiority, no narration, no explaining. Everything a character feels has to be inferable from what they do and from what they conspicuously avoid saying. That makes subtext the entire craft, and subtext is the first thing a model trained toward clarity and helpfulness destroys.
It also makes the genre problem acute. Crime, war, addiction, abuse of power: the standard subject matter of prestige drama sits directly on top of what mainstream assistants are trained to hedge. OpenRogue's lineup will write the interrogation scene, keep the antagonist's logic intact, and let a character be genuinely wrong without a paragraph of authorial correction attached.
The line on the page should be about the weather while the scene is about the divorce. Models default hard to on-the-nose dialogue because their training rewards being understood, which is exactly backwards for drama.
A screenplay is mostly dialogue, so any bleed between characters is immediately visible. Every speaker needs their own rhythm, vocabulary and evasions, sustained across a hundred pages.
Act breaks, turns, escalation, the midpoint that flips the goal. Diagnosing why act two sags is analytical work, and it benefits from a model that will reason deliberately rather than reassure you.
Most model-written scenes have a warm-up and a wind-down. Getting a model to cut both — start at the moment of conflict, end on the turn — takes instruction the model will not supply on its own.
Antagonists, bigots, cowards, and people whose reasoning is coherent and monstrous. A model that softens them cannot write half the roles worth acting.
Filtered models write characters who articulate their emotional state clearly and considerately, because that is the register they were tuned into. On screen it reads as therapy-speak, and it is the single most recognisable tell of AI-written dialogue.
Assistant training likes closure. Confrontations get defused, arguments end with mutual understanding, and the tension you spent three pages building evaporates in the last exchange. Drama requires scenes that end worse than they started.
Ask for a realistic interrogation, a plausible drug operation, or a hate-filled monologue that has to convince, and mainstream models hedge, abstract or decline. The refusals cluster exactly where prestige television lives.
Even without refusals, tone gets flattened toward the polite middle. Menace comes out as mild disapproval, class differences vanish from diction, and every character ends up speaking the same professionally neutral English.
| Model | Role | Why |
|---|---|---|
| Magnum v4 | Dialogue and scene work | Its training target was prose quality with restraint rather than volume, which translates directly into dialogue that trusts the audience. It will write the scene where the important thing goes unsaid, and it will not add a line explaining it afterwards. |
| Hermes 4 | Structural diagnosis | Turn Thinking mode on and it will actually reason through why your second act stalls — what the protagonist wants in each sequence, where the escalation flattens. Its neutral alignment means it argues a position instead of listing considerations. |
| UnslopNemo | Fast pass drafting | Fine-tuned against recycled AI phrasing, which matters more in dialogue than anywhere else — clichéd lines are unusable rather than merely dull. Fast enough to draft a dozen versions of a scene and throw eleven away. |
| Hermes 3 405B | Whole-script coherence | With 131K context it can hold a full feature and reason across it: setups without payoffs, a character who disappears for forty pages, a theme stated in act one and abandoned. That is the pass nobody wants to do by hand. |
OpenRogue is not screenwriting software. It does not enforce Fountain or industry formatting, does not paginate, does not track revisions or coloured pages, and does not export to Final Draft — you will be copying text into a real screenwriting app and formatting it there. Models are also weaker at screenplay format than at prose in general, since there is far less of it in training data, so expect to correct slug lines and parentheticals. And the fundamental limit stands: a model can generate a scene, diagnose a structure and imitate a register, but the reason the film should exist has to come from you.
/// Subtext
Rewrite this scene so neither character states what they actually want.
/// Structure
Think it through: why does my second act stall on page 62? Be specific.
/// Voice test
Names removed. Guess who says each line, and tell me which ones are ambiguous.
It can write scenes, draft alternatives quickly and diagnose structural problems in a way that genuinely saves time. It cannot originate the reason a story matters, and a full script generated end-to-end reads exactly like one. Treat it as a very fast collaborator on a draft you are steering, not as a writer.
Because the models were trained to be understood. Helpfulness training rewards clarity, and clarity is the opposite of subtext — so characters explain their feelings instead of concealing them. The fix is an explicit instruction to remove every stated emotion, plus a model like Magnum v4 that was tuned toward restraint rather than volume.
Yes. Crime, addiction, war, coercion and characters whose reasoning is coherent and repugnant are standard dramatic material, and OpenRogue adds no filter layer. The models will keep an antagonist's logic intact instead of undercutting it, which is what makes the role playable.
No. There is no screenplay formatting, pagination or export — you copy text out into whatever writing software you use. Models are also less fluent in screenplay format than in prose, so slug lines and parentheticals usually need a cleanup pass.
Hermes 3 405B, because 131K of context is enough to hold a feature and reason across the whole thing at once — unpaid setups, contradicted character logic, redundant scenes. Hermes 4 70B with Thinking mode is the better choice for outline-stage structural work.