Improve Your Screenplay With AI (Without Letting AI Write It)

You have a structured, formatted draft, and somewhere in it there is a problem you cannot see because you wrote it and have read it eleven times. The value of an AI consultant here is not intelligence. It is detachment. And because the assistant has read the whole project, you never paste anything in: you can just say "this scene".

Ask a question with no edges and you get advice like "deepen the character, raise the stakes, consider the pacing". Not wrong, just useless: you could have written it yourself before breakfast. Give the question edges instead: ask for the three weakest moments, by name, with reasons, and tell it not to suggest fixes yet. If you let it jump to solutions, it stops looking.

In SIGNAL that surfaced the line where Mara explains the observatory: exposition with no listener. It was right.

Being told something is weak is not the same as being told to change it. A moment can be quiet on purpose; a character can be unclear because the film is about not understanding them yet. Read the reason, and if it does not apply to the film you are making, move on without guilt. Of three notes, the lesson keeps two and refuses the one that wanted the ending explained.

Asking the assistant to argue against a choice changes the register of the reply. Without that phrase it is agreeable; with it, it pushes. Here it found that Mara stays in the station because the script needs her to, which is not a reason.

Then use it to think, not to write: ask for five options. The one chosen is small and physical: she reaches for her coat, then does not put it on. Entirely on screen, and it costs nothing to generate. The other four are dropped. Nothing has to be used just because it appeared.

The hammer is "rewrite this scene". You get a scene that is technically better and no longer yours. The scalpel names three things: the exact target (Mara's dialogue on page one), what must not change (every action line), and a request to see both versions so you compare rather than accept.

The new version is better: she explains nothing and says one word into a dead microphone out of habit. That is a character. The other half of this technique is asking what the audience could see instead. It turns a writing problem into a directing problem, and directing problems come with pictures, which puts you halfway into the storyboard.

Keep conversations separate: one for structure, one for the scene you are wrestling with, one for production questions. A single conversation that drifts across all three carries context that quietly bends every answer.

Open questions make the assistant reach for a new character or a new location, the cheapest surprise in prose and the most expensive in production. Take those tools away and it has to be cleverer. That is how SIGNAL got the reflection in the window, the image the final shot is built on for the rest of the course.

• Contradictions are cheap to fix in words and expensive in pictures. SIGNAL said the door had not been opened in 20 years, then put a chair against it. Somebody put that chair there. Audiences will not articulate that, but they will feel it.
• Pacing: ask it to point at the exact line where the scene drags. "Where does it drag?" gets you an essay; a line gets you a place you can fix.
• Tone: use it when you suspect you settled for the obvious version. Most of what comes back you will not use. Here, half a suggestion survived: the voice on the radio knows something about the room it should not know.

The assistant has read your project, not your intentions. "Is this ending good?" asks about an abstraction and gets a review. "This ending has to feel like a decision she makes, not something that happens to her. Does it?" asks about a target and gets a note.

Then ask what will be hard to produce. In SIGNAL it flagged Mara crossing the room quickly, because sustained full-body movement is where AI video falls apart, and suggested cutting to her hands instead. A better shot anyway, and now one that will survive generation.

And when a suggestion is good but not yours, say no. The assistant proposed making the radio voice ambiguous, possibly Mara's own. A good idea, and not this film. The whole consultation took eight messages and around 50 credits.

1. 1"What are the three weakest...?" A number forces a ranking, and a ranking forces a judgment.
2. 2"Argue against my version." Otherwise it agrees with you.
3. 3"Suggest five alternatives. No new characters, no new locations." Constraints make it work harder.
4. 4"Rewrite only this. Keep everything else. Show me both." The scalpel.

With a script you are happy with, Lesson 4 gives every character, place and object a description the image generation will read word for word.