Novel to Script With AI: What the Conversion Tools Get Right, What They Flatten, and a Workflow That Holds

What a converter does to a passage

Take a prose paragraph:

Mara did not answer. She looked at the receipt again, at the date, and thought about Lisbon, about the two days he had said he was in Leeds. She folded it once and put it in the pocket of her scrubs. "You should sleep," she said.

A naive AI conversion:

INT. HOSPITAL KITCHEN - NIGHT

Mara doesn't answer. She looks at the receipt again, at the date.

MARA (V.O.)
Lisbon. The two days he said he was in Leeds.

She folds it once and puts it in the pocket of her scrubs.

MARA
You should sleep.

Every element is formatted correctly and the scene is wrong. The interior thought became voiceover, the default move of every model asked to preserve content. The action lines are the prose sentences with tense changed. The scene has the novel's rhythm, not a film's. What an adapter writes instead:

INT. HOSPITAL KITCHEN - NIGHT

Mara reads the receipt. Reads the date.

She folds it once. Puts it in her scrubs pocket. Straightens the
blanket over him with the same care as before.

MARA
You should sleep.

Lisbon is gone from the page; it will come back in a later scene as an object or a line, when it is dramatic to know it. The care with the blanket carries what the thought carried. That is a decision, and the converter cannot make it, because it does not know that the Lisbon reveal belongs in scene 30.

The lesson: give the converter instructions that encode the decisions, or accept a transcript.

The conversion tools, as of September 2026

SparkLore. A project can hold prose and script; the prose-to-script conversion takes point of view, format (US screenplay or Korean drama S# format) and per-scene instructions, and the script exports Fountain, FDX, PDF and DOCX. Disclosure: this is our product.

Sudowrite. Publishes a guide to using its tools to convert a novel to a screenplay and markets its Muse model as trained on screenplays, but has no screenplay format mode or Fountain/FDX export in its documentation (Sudowrite blog). Conversion is done in prose tools and formatted elsewhere.

Laper. An AI-native screenwriting app; its idea-first flow accepts DOCX, PDF, TXT or MD reference files, so a manuscript can inform the generated outline and scenes, but there is no dedicated adaptation flow (Laper docs). Exports FDX and PDF; Fountain export "is not currently available" per its docs.

Squibler. Generates screenplays with "industry-standard formatting" from a description; no Fountain or FDX export found.

General chat models. Paste a chapter, ask for Fountain, get Fountain. No memory of the scene list, no bible unless you paste it, and the output needs a Fountain-aware editor to render. The cheapest route and the one that most needs the workflow below.

Arc Studio, WriterDuet, Final Draft. Screenwriting editors without in-editor AI drafting; the destination for a converted script, not the converter.

A workflow that holds

1. Adapt first, on paper

Before any conversion: choose the spine (whose story, what they want, what it costs), build a scene inventory from the novel, and mark each chapter's material as scene, merged, summary or cut. The scene inventory method is the full procedure. The output is a scene list of 40 to 70 scenes for a feature, or an episode-by-episode list for a series, each with location, characters, goal and outcome.

Converting chapter by chapter in the novel's order is the mistake that produces a 300-page script with the novel's pacing. The scene list is the corrective.

2. Put the cast and world in a bible

Names, ages as they will appear in cues, locations with the slugline you will use for each (INT. HALE HOUSE - KITCHEN, not three variants), and the rules of the world in one line each. Every conversion prompt reads from this, so cues and sluglines come out consistent. Our story bible template is the structure; the format guide has the slugline conventions.

3. Convert one scene from the list, with decisions encoded

Source: novel chapter 7, paragraphs 12–31 (pasted). Scene 18 on the list.
Format: US screenplay, Fountain syntax.
Slugline: INT. HALE HOUSE - KITCHEN - NIGHT.
Characters present: MARA, HALE. (Cues from the bible.)
Goal: Mara learns he lied about Leeds. Outcome: she says nothing.
Rules:
- No voiceover, no flashback. Interior thoughts become action or are cut.
- Keep at most half the novel's dialogue; convert the rest to behaviour.
- Do not reveal Lisbon; that is scene 30.
- Action lines: present tense, one action per line, no camera directions,
  no adjectives about emotion.
- End on an action, not a line.
Length: about 1.5 pages.

The rules section is the adaptation. Without "no voiceover" you get voiceover. Without "do not reveal Lisbon" you get the novel's information at the novel's moment. Without "half the dialogue" you get every line, because dialogue is the part of prose a model treats as already finished.

4. Read, fix, then use it as context for the next

Fix the scene before converting the next, because the converter has no memory beyond what you paste. Attach the fixed scene 18 to the prompt for scene 19 so tone and cue names carry. In a tool with a project, the bible does this automatically; in a chat model, the paste is the memory.

5. Structural pass on the whole

When the scene list is converted, read the script alone, without the novel. Watch for:

• Novel pacing. Long scenes early, rushed scenes late. Scripts front-load less.
• Exposition that survived. The model keeps facts; a script needs them only when dramatic.
• Interior leakage. Any "(thinking)" parenthetical, any character saying what they feel, any action line that reports an emotion ("Mara, devastated, ...").
• Cue drift. MARA in one scene, NURSE MARA in another. The bible should have prevented it; check anyway.
• Length. Feature: 90 to 120 pages. If you are at 160, the scene list was too long, not the scenes.

Ask for a diagnostic pass: "List every action line that describes a feeling rather than a visible action, with scene numbers." Then fix by hand.

6. Format and export

If you converted in a chat model, the output is Fountain text; open it in a Fountain-aware editor to render pages, or import it into your screenwriting app. If you converted in a tool with export, produce a Fountain or FDX master and a PDF for readers. The format comparison explains which to keep; the short version is: keep Fountain or FDX, send PDF.

Korean drama adaptation

Web novels adapted into Korean dramas follow the same workflow with a different format target: S# scene numbering, prose stage directions, name-colon dialogue and the (E)/(N) tag system rather than the American column layout, and an episode structure of typically 16 episodes for broadcast or 8 to 12 for streaming. The conversion prompt changes its format line; the scene list becomes an episode-by-episode list with an ending hook per episode. The K-drama format guide has the conventions, and the 16-episode structure post covers the arc. SparkLore's converter and free script generator both offer the Korean format as a target, and the drama script formatter fixes alignment for pasting into HWP or Word.

Rights

Your own novel: proceed. Someone else's: adaptation is a derivative work and needs the rights, whether or not AI is involved. If the novel is under a publishing contract, check whether dramatic rights were granted.

Try it in SparkLore

SparkLore keeps the novel, the scene list and the script in one project with a shared bible, converts prose to script in US or Korean format with the decisions above as settings, and exports Fountain, FDX, PDF and DOCX. Without an account, the AI script generator converts one scene description into formatted pages in either format, and the prompt library has the "novel paragraph to drama script scene" prompt as a form. Sign up to import a manuscript and start the scene list; prices are on the pricing page.