# AI Script Breakdown: What It Gets Right and What to Check

Source: https://storiara.com/blog/ai-script-breakdown
Last updated: 2026-09-14
Author: Nick Harty, Storiara

> AI script breakdown tools read a screenplay and pull out scenes, characters, locations, and named props, wardrobe, and effects in minutes. They do well with anything written explicitly and badly with what's implied: multiples, counts of extras, stunts hidden in verbs, wardrobe continuity by script day, and department ownership. Treat the output as a fast first pass and review every scene before it drives a schedule or budget.

## How these tools work, roughly

Most AI breakdown tools run in two stages, whether they say so or not.

The first stage is structure. Scene headings in screenplay format follow strict conventions (INT. or EXT., location, time of day), so splitting a script into scenes and reading INT/EXT and DAY/NIGHT can be done with ordinary parsing rules. Page eighths can be measured from where headings sit on the PDF page. None of this needs a language model, and when it's done with rules it tends to be reliable on cleanly formatted scripts.

The second stage is extraction. A large language model reads each scene's text and returns lists: characters, props, wardrobe, vehicles, effects. This is where the speed comes from and where the errors live, because the model has to decide what a sentence means.

Knowing which stage produced an element tells you how much to trust it.

## What AI extraction gets right

On a well-formatted script, these usually come out clean:

| Element | Why it works |
|---|---|
| Scene list, INT/EXT, DAY/NIGHT | Fixed format in the heading |
| Speaking characters | Character cues are formatted and capitalized |
| Locations and sets | Named in headings |
| Explicitly named props | "She loads the revolver" names the revolver |
| Named vehicles and animals | "His '87 Volvo", "the lab, Biscuit" |
| Obvious effects | "Rain hammers the roof", "the car explodes" |
| Named wardrobe | "a yellow slip dress" |

That's a large share of a breakdown's raw inventory, and it's the tedious part. On a 100-page script, having scenes, characters, and named props listed before you start saves real typing time, and it catches things a tired human skips on page 87.

## What it gets wrong or misses

Here's the same kind of list, for the categories that need a human:

| Problem | Example | Why AI struggles |
|---|---|---|
| Implied items | "Maya reads the text" (phone never named) | Nothing to extract; it requires inference about the physical world |
| Multiples | A bottle smashed over a head needs breakaways, several | The script says one bottle |
| Background counts | "The bar is packed" | No number exists in the text |
| Stunts hidden in verbs | "She slips on the wet ramp" | The risk is never stated |
| Practical versus VFX | "The car explodes" | The script can't answer it; the director and departments decide |
| Wardrobe continuity | Same clothes as the previous scene, second day | Requires tracking script days across scenes |
| Name variants | "the pickup", "Dad's truck", "the Ford" | May become three vehicles |
| Department ownership | A lamp an actor switches on (prop) versus a lamp nobody touches (set dressing) | Depends on the action, and conventions vary by show |
| Invented items | A "gun" listed in a scene where a character only mentions one | Models sometimes list things that are discussed, remembered, or imagined |
| Minors and sensitive work | A 16-year-old clerk; an intimate scene | Needs flags for permits, teachers, coordinators |

Name variants are an old problem, and not unique to AI. Final Draft's manual shows its Tag All command catching "gold key" but not the same object when the script just says "key," so you tag the variant and point it at the existing tag. A language model can sometimes merge variants better than exact matching can, and it can also merge two things that are actually different, like the hero police cruiser and three parked dressing cars.

## One scene, compared

Scene 12 from a fictional feature:

```
12  EXT. RIDGEVIEW HIGH - PARKING LOT - NIGHT                 12

    Homecoming. The lot is jammed. JUNIE (17) sits on the
    hood of her brother's Civic, texting. Bass thumps from
    the gym.

    A SECURITY GUARD sweeps his flashlight across the cars.
    Junie slides off the hood, ducks between two pickups,
    and her dress catches on a trailer hitch and rips.

                         JUNIE
                    (into phone)
               Tell Marco I'm not coming in.

    Fireworks burst over the football field.
```

Below is an illustrative extraction, built from the error types in the table above, next to the reviewed breakdown:

| Category | Illustrative AI output | After human review | What changed |
|---|---|---|---|
| Cast | Junie, Security Guard, Marco | 4. Junie; guard as silent bit | Marco is only mentioned in dialogue; the guard has no line |
| Background | (none) | 40 students and parents arriving in the lot, homecoming dress | "The lot is jammed" became a count |
| Vehicles | Civic | Junie's brother's Civic (hero); 2 pickups, one with trailer hitch; 25 to 30 parked cars (dressing) | Parked cars fill the frame too |
| Props | Phone, flashlight | Junie's phone (practical screen or replacement); guard's flashlight | Screen plan noted |
| Wardrobe | Dress | Junie: homecoming dress, clean and ripped versions, multiples for resets | Destroyed in scene |
| SFX / VFX | Fireworks | Fireworks: practical pyro with permits, or VFX; decide | Open decision written down |
| Stunts | (none) | Junie slides off a hood and ducks between trucks at night: coordinator to assess | Hidden in verbs |
| Sound | Bass from gym | Music playback from the gym, cleared track or temp | Rights question |
| Notes | (none) | Minor performer at night: hours and permits; hood of a car (protect paint) | Flags added |

In this illustration, about half the lines were usable as written and the rest needed a count, a multiple, a safety flag, or a reclassification. Real results vary by tool and by script, so test before committing, and plan on a review pass either way.

## How to test a tool before you trust it

Pick three scenes you've already broken down by hand: one dialogue scene, one with background, one with action. Run them through the tool and score each category:

1. Count elements the tool found that your breakdown also has.
2. Count elements your breakdown has that the tool missed.
3. Count elements the tool listed that shouldn't be there.
4. Note anything that landed in the wrong category.

Misses in cast and locations are serious, because those drive the schedule. Misses in props are annoying but get caught in department review. Invented stunts or weapons are worth knowing about early, because they put the wrong people on call sheets.

## The review pass, scene by scene

For every scene, after extraction:

- Read the action lines for verbs involving risk and add stunts.
- Put a number on every background group.
- Add implied items (phones, keys, bags, leashes).
- Add multiples for anything destroyed, spilled, wet, or bloodied.
- Check wardrobe against the script day and the previous scene.
- Merge name variants and split anything merged by mistake.
- Decide practical or VFX for every effect, or write the open question.
- Flag minors, animals, weapons, intimacy, water, heights, and night exteriors.
- Delete anything the script only mentions in dialogue or memory.

This is the same judgment a paper breakdown needs. The difference is you start from a filled-in list instead of a blank sheet. Our [script breakdown example](https://storiara.com/blog/script-breakdown-example) shows the full human version of a scene.

## How Storiara's extraction works

Storiara splits scenes deterministically from the script's headings (INT, EXT, EST, I/E, including numbered headings) and computes each scene's page count in eighths from the PDF layout. It then uses Google Gemini models to extract locations, characters and extras, props and wardrobe, and music, effects, and stunts. Each AI-extracted item is checked against a quote from the script. That check depends on a quote existing, so implied items still have to be added by a person. Results appear as extraction runs, failed extractions retry, and you can use Retry Extraction on the Scenes page.

For the review pass, the Script & Breakdown View lets you highlight script text and add it as a character, prop, wardrobe item, vehicle, stunt, extra, effect, and other categories, delete wrong highlights, and hide terms you don't want detected again. Items you add by hand are preserved when content is re-pulled from the script. Storiara doesn't build the schedule for you; scenes go on a drag-and-drop stripboard you arrange. Script text is sent to Google's Gemini API for processing, so check your NDA before uploading. See [script analysis](https://storiara.com/feature/script-analysis) and [tagging breakdown elements](https://storiara.com/docs/how-to/tag-breakdown-elements).

## Where AI breakdown is worth it

It pays off most on long scripts, on productions that revise often (re-extracting changed scenes beats retyping them), and for people who haven't done a breakdown before and need to see the shape of one. It helps least on short, effects-heavy scenes where nearly every line is a judgment call. For side-by-side tool comparisons, see [best script breakdown software](https://storiara.com/blog/best-script-breakdown-software).

## Frequently asked questions

### Can AI do a script breakdown?

It can do a useful first pass. Current tools extract scenes, speaking characters, locations, and explicitly named props and effects quickly. A person still has to add implied elements, counts, multiples, and safety flags, and check for items the model got wrong.

### How accurate is AI script breakdown?

It depends on the tool, the script, and the category. Explicit nouns like a named prop or a scene heading come out well. Implied items, background counts, and continuity across scenes are where errors concentrate. Test any tool on a few scenes you've already broken down by hand.

### Will AI replace the 1st AD's breakdown?

It moves the work. Instead of typing every element, the AD or UPM reviews and corrects extracted lists, then adds the judgment calls: what's a stunt, what needs multiples, what's practical versus VFX, and what the director actually wants on camera.

### Is it safe to upload a script to an AI breakdown tool?

Read the tool's terms and privacy policy, and check your NDA or the studio's security rules first. Many tools send script text to a third-party AI model provider, so find out which one and under what terms before uploading unreleased material.

## Sources

- [Final Draft 13 User Guide (Mac), Tagging an Item (exact-match Tag All example)](https://www.finaldraft.com/downloads/manuals/fd13-user-guide-mac-exported.pdf)