What “AI screenplay formatting” usually gets wrong
· The StoryProp team
Ask a general-purpose chatbot to format a scene as a screenplay and you usually get something that passes a squint test: the heading in capitals, the character name near the middle, the dialogue indented more or less correctly. Then you paste it into the file where the script actually lives, add a line of action on page nine, and the whole thing drifts — cues wander into the dialogue column, a parenthetical becomes its own paragraph, and the page numbers stop meaning anything. What went wrong is not capacity but construction: you were handed text that looks like a screenplay and was never built like one. Size is not the obstacle: a feature runs 80 to 125 pages in 12-point Courier under the Academy Nicholl Fellowships entry criteria for the 2026 cycle, and a script that size is no longer a large document for a current model. The model can hold your entire screenplay. It formats the surface of one.
The mistake underneath is older than any AI feature, and most screenplay format software has inherited it at one time or another: treating format as appearance. On that view, formatting is a set of indents you apply to text, and a formatter is a machine that applies them. It is also backwards. In a screenplay the margins are not the format; they are the visible residue of a decision about what each line is.
Margins are outputs, not inputs
Every line in a screenplay belongs to a type: scene heading, action, character cue, parenthetical, dialogue, transition. Dialogue occupies a narrow column in the middle of the page not because anyone liked the look of it but because it is dialogue, and that is where dialogue goes. The indentation is generated from the type. Run the relationship the other way and you get a document that remembers geometry and nothing else — a file that knows exactly where a line begins and has no idea why.
The difference stays invisible until you ask the document to do something. Restyle a script to a different template and a semantic file re-derives every margin, while a geometric one has to be corrected paragraph by paragraph. Ask which scenes Daniel speaks in and the semantic file answers, because it knows a character cue when it is holding one. Search the geometric file and you are really searching for a name in capital letters, which will also turn up the line where somebody SLAMS the door.
That is what template-thinking costs. A tool that formats by inserting tabs and capitals can produce a page that is correct at this moment, but the meaning was thrown away when the page was made. Everything downstream — pagination, revision, export, even a count of dialogue against action — has to reconstruct that meaning by pattern-matching, and pattern-matching is right most of the time. Most of the time is a poor standard for a document a reader will treat as evidence that you know the job.
The grammar under the page
The element list is short, and it behaves like a grammar: the types have rules about what may follow what. A parenthetical exists only inside a dialogue block. A character cue is always followed by dialogue, or by a parenthetical and then dialogue. A transition closes a sequence. Dual dialogue is two blocks running side by side, which is a structural fact before it is a layout. A tool that holds the grammar can enforce it as you type; a tool that holds only indents can only hope.
Take a stretch of the demo manuscript we keep around here, The Quiet House. A scene heading — INT. KITCHEN - NIGHT — tells the reader where and when, and tells the file that a new scene has started, which is what makes scene numbering and scene-level navigation possible at all. A line of action follows: Mara rinses the chipped cup, sets it on the drainer, turns it so the chip faces the wall. Then a character cue, DANIEL, then a parenthetical, then his line. Five lines, five element types, and every margin on the page falls out of them.
Guessing fails in specific, repeatable places. A line of action written in capitals for emphasis is promoted to a character cue. A heading that does not open with INT. or EXT. — LATER - THE SAME KITCHEN, say — is demoted to action, and the scene quietly vanishes from the scene list. A beat typed in parentheses inside an action paragraph turns into a parenthetical attached to nothing. Extensions such as (V.O.), (O.S.) and (CONT'D) belong to the cue, and a parser that reads them as parentheticals splits the block in two.
Pagination is a calculation, not a layout
Pages in a screenplay are computed, continuously, from a fixed-pitch face at a fixed size and a fixed number of lines to the page. Courier at 12 point is the face the Academy names in the Nicholl entry criteria, and naming the face is what makes a page count comparable from one script to the next. Add three words to a speech on page twelve and everything after page twelve moves. A tool that formats once has produced a photograph of a document that was briefly correct.
Real pagination also enforces small courtesies that readers notice only when they are missing. A page does not end on a character cue with the speech beginning overleaf. A speech split across a break takes MORE at the foot of the page and the character's name with CONT'D at the top of the next. A scene heading stranded at the bottom of a page moves down. Dialogue breaks at a sentence boundary where one is available. None of it can be faked with blank lines, because the next edit moves the break.
Writers feel this as structure. Length is a signal: a scene that runs four pages when you thought it ran two tells you something before any note-giver does. The familiar one-page-per-minute conversion is a working rule of thumb from screenwriting practice rather than a figure any guild or the Academy publishes, so treat it as a rough gauge. The page band, though, is published and is enforced: the Nicholl entry criteria for the 2026 cycle set a hard minimum of 80 pages and a hard maximum of 125, tightened from the looser 70–160 range that older rules summaries still quote, where 80–125 was suggested rather than required. If your pagination is decorative, you are guessing at a number a reader will check.
| Rules | Accepted length | What the rules add |
|---|---|---|
| 2026 cycle entry criteria | 80–125 pages | Hard minimum and hard maximum; 12-point Courier, standard industry format |
| Earlier rules summaries (2023/2024 era) | 70–160 pages | 80–125 pages suggested rather than required |
Academy of Motion Picture Arts and Sciences, Nicholl Fellowships in Screenwriting entry criteria, 2026 cycle, with earlier rules summaries for comparison
What survives a round trip
Two interchange formats matter in practice. FDX is XML: element types are stated outright, alongside a production layer of revisions, scene numbers, notes, and page locks. Fountain is plain text with conventions — an uppercase line followed by dialogue is a cue, a leading period forces a scene heading, a leading angle bracket forces a transition, a caret marks dual dialogue. Both are explicit formats, in the sense that what each one carries is written down — which is what makes it possible to say precisely what survives a trip through them.
It helps to sort a script into three layers before exporting anything. The content layer — the sequence of elements and their text — travels reliably. The presentation layer — pagination, MORE and CONT'D, running headers, the exact margins — is regenerated at the destination from that application's template, and should be expected to differ. The production layer — revision marks and colored pages, locked pages and A-pages, scene numbers, script notes — is where round trips actually break, because plain text has nowhere to keep most of it.
So test it once, on something that does not matter yet. Build a page holding every awkward case you can think of: dual dialogue, a forced scene heading, a cue with an extension, a transition, a note, a scene number, and a speech long enough to break across pages. Export it, open it in the other application, export it back, and read the two versions side by side. One pass of that tells you exactly which layer survives the trip and which one is rebuilt at the far end.
Fountain deserves a note of its own, because its trade is easy to misread. Being a convention over plain text, a Fountain file opens in any editor on any machine and will keep opening long after any particular application has changed hands, which is a real kind of safety. The cost is that anything the convention does not name has to be smuggled in as a note or left behind. For drafting, outlining, and moving pages between machines, that is a good trade. For a production draft carrying revision colors, it is the wrong container.
Four questions to ask of any formatter
Four questions do most of the work. Does the document store element types, or only indentation? Does pagination update while I write, or only when I ask for a preview? What happens to dual dialogue, a forced heading, and my notes on the way out and back in? Can I get FDX and Fountain out of it directly, without a detour through a word processor? And if the vocabulary is still new, the elements and how they behave are worth reading before anything else — the grammar is what the machinery exists to hold.
How StoryProp is built
In StoryProp's screenwriting work, screenplay elements are semantic, so layout is produced from them rather than typed on top of them; pagination is authoritative in preview and export; and FDX and Fountain are there for the moment a script has to move. The effect while drafting is that revision does not degrade the file. You point at a passage, say what you want changed, and what comes back is still a correctly typed sequence of elements — the rest of the scene is left alone, not quietly reflowed into prose wearing a screenplay costume.
A document that knows what its lines are can tell a cue from a shout, recompute a break after the fifth pass, and hand the whole script to another application with its layers intact. That is worth setting up once. Format is the least interesting thing about a screenplay, and the reward for getting the machinery right is that it stops being something you think about at all: the page count means what it says, the margins take care of themselves, and the whole of your attention goes back to the scene.
If you would rather write scenes than nudge margins, StoryProp keeps screenplay elements as elements — cues, parentheticals and dialogue hold their types through every revision, layout and pagination follow from them, and FDX or Fountain is there when the script has to travel.
Sources
- Feature screenplay page limits and 12-point Courier requirement — Academy of Motion Picture Arts and Sciences, Nicholl Fellowships in Screenwriting entry criteria, 2026 cycle, with earlier rules summaries for comparison
Rates, fees and market figures change. These were accurate at the dates shown; check the source for current numbers before you rely on one.