Writing Story & Dialogue for AI Video

A premise is the situation ("two families clash over wedding-day logistics"). An idea is what the film is about ("everyone wants to be the main character at someone else's celebration"). A story is someone wanting something, being obstructed, and something changing by the end.

Under 60 seconds, there's no room to build a real arc. Pick one shape and commit to it.

Don't mix two. Nearly every flat short film is an escalation that abandons itself for a demonstration halfway through.

The turn has to be planted in the setup — inevitable in hindsight, unexpected in the moment. And the button is an image first: if the ending only works because someone says the line out loud, it's an announcement, not an ending.

Score each axis 1–5. Anything at 3 or below is what to fix next — no amount of production value repairs a story problem, and this is the cheapest possible moment to catch one.

Models write competent, complete, explanatory sentences by default. Real dialogue — especially comic dialogue — is none of those things.

• Cut the first sentence. Almost every written line begins with a run-up. The real line usually starts one sentence later than you wrote it.
• Put the punch word last. The final word is the one that lands — rewrite the clause order until the sharpest word is at the end, then stop.
• Never explain the joke. If the visual is already doing the work, the line shouldn't mention it too.
• Make one noun specific. A named, specific detail is funny (or moving); a generic one isn't. Specificity is the single highest-yield edit on any line, in any genre.
• Shorter than feels safe. Cut until removing one more word would break the sense. Then cut one more and check whether it actually broke.
• Let someone not answer. The strongest reply is often a look, a pause, or a reply to something else entirely.
• Give each character a different register. If you could swap a line between two characters without anyone noticing, you've written one voice, not two.
• Subtext: say the small thing, mean the big one. Nobody announces their own anxiety out loud; they mention something small instead.

Creator-surface dialogue is not just shorter ad copy. It needs to sound like a person discovered something worth saying inside a real moment. The first line should feel sayable in one breath, and it should carry a point of view or a lived problem before it carries a sell.

The single most reliable comic structure: one character makes a claim; the other doesn't argue with it — they escalate it absurdly, or deflate it flatly, in a line shorter than the first. Almost every memorable comic exchange is built this way. Two people bragging past each other for twenty seconds, never actually answering one another, reads as two monologues, not a scene — the fix is making the second line a direct response to the first, not a parallel complaint.

A plan bullet is not a locked line. Roles like "phone thought" or "end line energy" are concepts — not dialogue you may generate or ship as if the client approved them.

• Grep the brief for the verbatim string before producing any VO, SPEAK, or dub line.
• If only a concept exists, say so immediately, draft the line as a labelled RECOMMENDATION, and carry an approval flag until the client answers.
• Never promote a plan bullet to canon by acting on it.
• Keep a name glossary for ambiguous labels (internal beat ID vs client word) so ASL names never collide with logo-card language.

These are technical constraints, not style notes — ignore them and a good line will still fail in production:

• One speaker per generated clip. Two speakers in one generation call causes line bleed and swapped delivery.
• One line per clip, roughly 8 seconds of speech maximum. Longer lines get rushed unnaturally.
• Under about 10 words where lip-sync matters. Short lines survive sync; long ones drift.
• Quote the line exactly, and make sure nothing else in the prompt reads like speech.
• Never leave audio on with no line and two people in frame. The model will invent speech to match the mouths — either give it an exact line, or turn audio off as a parameter, not a sentence.
• Write what the mouth does after the line ends. An unspecified mouth keeps moving past the end of the line.

The best line is often no line at all.

• A look beats a sentence in almost every beat that's meant to land.
• Overlap — someone starting to answer before the other finishes — reads as real; clean turn-taking reads as a script.
• Silence is a line. Write the pause into the shot list so the edit protects it, or it'll get trimmed away.
• Let the picture carry the plot and the dialogue carry the character. Dialogue that carries plot becomes exposition, and exposition kills a scene.

Each fix gets more expensive the later you catch it:

A story problem discovered after generation costs the entire shot. Discovered here, it costs a sentence.