AI Prompts for Screenwriting: A Teardown of a Bad Prompt

Write a tense scene between two characters, a detective and a suspect, in an
interrogation room. Make it dramatic and gripping.

Most screenwriting-AI guides are wrong about the same thing: they treat a screenplay like content to be generated in bulk. Feed the model a logline, get pages, repeat. But a script isn't prose you skim — it's a blueprint where every line is spoken aloud by an actor and watched by a room full of people who will notice the instant it rings false. Volume is the enemy. Precision is the point.

So instead of another list of "50 amazing ai prompts for screenwriting," let's take one weak prompt apart and rebuild it, so you can see the machinery.

Before: The Weak Prompt

Here's a prompt of the kind you'll find everywhere:

Write a tense scene between two characters, a detective and a suspect, in an interrogation room. Make it dramatic and gripping.

Looks fine. It'll produce pages. And the pages will be terrible in a specific, instructive way.

Why It Fails

Run that prompt and you get a scene where two people say exactly what they mean, loudly. "You did it, didn't you?!" "I'll never confess!" It's dramatic in the way a soap opera is dramatic — all surface, no pressure underneath.

It fails for four reasons.

No goal conflict. "Tense" and "dramatic" are effects, not causes. Tension comes from two people wanting incompatible things. The prompt never says what either character wants, so the model fakes tension with volume and exclamation points.

No subtext. Great screen dialogue is people not saying what they mean. The prompt asks for drama, so the model makes everyone say the quiet part loud. Subtext has to be requested explicitly or it evaporates.

No format discipline. The model doesn't know if you want proper screenplay format, so it guesses — usually a messy hybrid of prose and script that no reader will take seriously.

No constraint. With nothing to push against, the model reaches for the most familiar version of "interrogation scene" it has ever seen. You get the cliché because you asked for the category, not the specific.

⚠️ Common mistake: Asking for the feeling you want ("make it tense," "make it emotional") instead of the mechanism that produces it (opposing goals, a secret, a ticking clock). Models can't manufacture emotion on command, but they can execute a mechanism. Name the mechanism.

After: The Improved Prompt

Here's the rebuild:

Write an interrogation scene in standard screenplay format. - DETECTIVE wants a confession before her shift ends in 10 minutes; she's bluffing, she has no real evidence. - SUSPECT is innocent of THIS crime but guilty of something else he's desperate to hide. Neither states their real goal. The scene is a game of who blinks first. No character says exactly what they mean. End on the suspect ALMOST slipping. Max one page.

What this does: It hands the model a working machine — competing goals, a secret on each side, a time pressure, a ban on on-the-nose dialogue, a format, and a length cap. Now the model has something to execute instead of a mood to fake.

Breaking Down Each Element

The opposing goals ("confession before shift ends" vs. "hide the other thing") create real friction. The detective's bluff is the engine — she's the one under pressure, which is a nice inversion of the cliché.

The secret asymmetry — he's innocent of this but guilty of that — is what generates subtext automatically. He can't relax, but he also can't explain why, so every answer is a tightrope. You didn't have to ask for subtext; you built a situation that requires it.

The time pressure ("10 minutes") gives the scene a clock. Clocks make everything sharper.

The bans and caps ("no one says what they mean," "max one page") stop the model's two worst instincts: going literal and going long.

⚡ Pro tip: The phrase "end on X almost happening" is worth memorizing. "Almost" is where screen tension lives — the near-confession, the near-kiss, the near-miss. Asking for the almost, rather than the event, gets you scenes with a charge instead of scenes with a resolution.

Variations for Different Contexts

For a TV writer breaking a comedy scene: swap the mechanism. "Both characters want to end this conversation but politeness traps them; each tries to make the other leave first." Comedy is just conflict with lower stakes and worse solutions.

For an ad creative writing a 30-second spot: "Two lines of dialogue max. The product solves a problem neither character will admit they have." Constraints get tighter as the runtime shrinks.

For a game writer doing branching dialogue: "Write this exchange three ways — one where the player was kind earlier, one cruel, one neutral. The NPC remembers." The variation is the deliverable.

⚡ Pro tip: Always ask for proper screenplay format explicitly, and paste one page of correctly formatted script as an example. Models are inconsistent with sluglines and parentheticals unless you show them exactly what you mean. An example is worth three sentences of instruction.

Testing a Whole Sequence, Not Just One Scene

Individual scenes can each be strong while the sequence between them drags. Screenwriters feel this as a script that "reads slow" despite good pages. A sequence-level prompt catches it:

Here are one-line descriptions of my next 6 scenes. For each, tell me: what does the audience LEARN, and what new QUESTION does it raise? If a scene teaches nothing new and raises no question, flag it as a candidate to cut or combine. [SCENE LIST]

What this does: It applies the "learn something / ask something" test to every scene. A scene that does neither is dead weight, no matter how well-written the dialogue is.

⚡ Pro tip: Great scripts rarely resolve a question in the same scene they raise it — they overlap. Ask the model to check whether your questions and answers are stacked (boring, predictable) or braided across scenes (propulsive). Braiding is why some scripts feel impossible to stop reading.

Punching Up Flat Dialogue

When dialogue reads flat, the fix usually isn't wittier lines — it's giving each character something to protect. Try this:

This exchange feels flat. For each character, invent one thing they're trying NOT to reveal in this conversation. Then tell me which of their existing lines would change if that secret were pressing on them. Don't rewrite — just point.

What this does: Flat dialogue is usually characters with nothing at stake in the words. Hidden agendas put pressure behind every line, and the model shows you where that pressure would surface.

⚡ Pro tip: The best on-screen lines often answer a question that wasn't asked, or dodge the one that was. Ask the model to flag any line where a character responds too directly — direct answers are where scenes go slack.

One More Format Warning

Readers in the film industry reject scripts on page one for format sloppiness, before a word of story registers. If you're using AI anywhere in your process, run a final format-only pass:

Check ONLY formatting in this page: sluglines, character cues, parentheticals, action-line length. List every deviation from standard screenplay format. Ignore the story entirely.

What this does: It separates format from content so nothing slips through. A brilliant scene in broken format still gets tossed on the coverage pile.

Save and Reuse This

The improved prompt above is a template with four slots: two goals, a secret, a clock, a format. Fill the slots differently and you can generate a heist scene, a breakup, a job interview from hell. The structure travels.

That reusability is exactly why saving prompts beats rewriting them. Drop this teardown template into PromptABCD, mark the fill-in-the-blank slots, and next time you're staring at a blank scene you start from a working machine instead of "make it dramatic." The difference between a weak prompt and a strong one is mostly the difference between a mood and a mechanism — save the mechanisms.