66%
The final tenth is more action-heavy
The median rises from 54% action around the midpoint to 66% at the ending. Dialogue is the complement.
Screenplay Anatomy Atlas · August 2026
The shape of a screenplay, measured
We normalized 2,077 structured screenplays from first scene to last, then measured what changed: action and dialogue balance, scene rhythm, speaking-character arrivals, and who carries the dialogue. These are corpus patterns—not instructions for where your story must turn.
What is actually in here? The included files range from Alien, Chinatown, Fargo, Inception, Moonlight, Parasite, and Whiplash to more than 2,000 other scripts. These are archived screenplay versions—not claims about the final cut or shooting draft.
2,200 records inspected2,077 includedNo one title dominates the resultNo screenplay text republished
Three findings first
The median is useful only when the variation stays visible
Every headline below is paired with its distribution later in the atlas. None of the figures identifies a “correct” screenplay or separates produced films from unproduced drafts.
65%
The speaking cast keeps arriving
Only 22% of the eventual speaking cast appears in the opening tenth; the median reaches 65% by halfway.
−0.523
Bigger casts spread speaking time wider
As the number of speaking characters grows, the three busiest voices usually account for less of the dialogue. The −0.523 score measures how consistently those two things move in opposite directions; it is not a quality score.
Movement across the draft
Openings and endings do not resemble the middle
Picture each screenplay as ten equal stretches of story text. Action accounts for a median 63.8% in the first stretch, settles near 54.5% around halfway, and reaches 66.1% in the final stretch. The shaded range stays wide: many scripts depart sharply from this contour.
Action share across the script
We split every script into ten equal stretches of story text. The line is the middle script; the shaded band contains the middle half.
Line: median · Shading: 25th–75th percentile
Choose a range, or tap or hover the chart, to inspect exact values.
View exact-value table
| Progress | 25th percentile | Median | 75th percentile |
|---|---|---|---|
| 0–10% | 52.6% | 63.8% | 74.3% |
| 10–20% | 43.9% | 55.7% | 66.0% |
| 20–30% | 44.7% | 55.1% | 66.6% |
| 30–40% | 43.6% | 54.7% | 65.8% |
| 40–50% | 43.1% | 54.5% | 65.3% |
| 50–60% | 44.4% | 54.5% | 66.2% |
| 60–70% | 44.5% | 55.7% | 67.3% |
| 70–80% | 44.5% | 56.3% | 68.5% |
| 80–90% | 47.3% | 59.9% | 73.3% |
| 90–100% | 54.0% | 66.1% | 78.5% |
Scenes within each tenth
Each bar counts how many scenes fit inside an equal amount of story text. A taller bar means shorter scenes on average—not a longer section.
Bars: median · Whiskers: 25th–75th percentile
Choose a range, or tap or hover the chart, to inspect exact values.
View exact-value table
| Progress | 25th percentile | Median | 75th percentile |
|---|---|---|---|
| 0–10% | 8 | 13 | 17 |
| 10–20% | 7 | 11 | 16 |
| 20–30% | 8 | 11 | 16 |
| 30–40% | 8 | 12 | 17 |
| 40–50% | 8 | 12 | 17 |
| 50–60% | 8 | 12 | 17 |
| 60–70% | 8 | 12 | 18 |
| 70–80% | 9 | 13 | 18 |
| 80–90% | 9 | 14 | 20 |
| 90–100% | 8 | 14 | 20 |
Scene count supplies a second signal. Each bucket contains the same share of narrative words, yet the median ending contains 14 scenes versus 12 near the midpoint. In this representation, the ending is composed of more, shorter scenes. It does not establish faster screen time or editing pace, which require the finished film.
Character arrival
New voices keep entering long after the opening
Divide a script into ten equal stretches of story text. In the middle script, only 21.7% of everyone who will eventually speak has appeared after the first stretch. That reaches 65.0% by halfway and 94.4% by the end of the ninth.
Supporting and one-line roles count, so this does not mean protagonists arrive late. It means a blanket “introduce the whole cast immediately” rule does not describe what these scripts actually do.
When speaking characters first appear
By each tenth, what share of everyone who will eventually speak has appeared? One-line roles count too.
Line: median · Shading: 25th–75th percentile
Choose a range, or tap or hover the chart, to inspect exact values.
View exact-value table
| Progress | 25th percentile | Median | 75th percentile |
|---|---|---|---|
| 0–10% | 16.1% | 21.7% | 28.1% |
| 10–20% | 28.8% | 35.7% | 44.4% |
| 20–30% | 38.8% | 46.5% | 55.6% |
| 30–40% | 48.5% | 56.0% | 64.6% |
| 40–50% | 57.3% | 65.0% | 73.1% |
| 50–60% | 65.9% | 73.4% | 80.3% |
| 60–70% | 74.1% | 80.6% | 86.6% |
| 70–80% | 82.5% | 87.7% | 92.5% |
| 80–90% | 90.9% | 94.4% | 97.4% |
| 90–100% | 100.0% | 100.0% | 100.0% |
Dialogue center of gravity
Three characters carry most dialogue—until the cast gets large
In the middle screenplay, the three most talkative characters deliver 57.2% of all dialogue. In scripts with fewer than 20 speaking characters, they carry about 72%; in scripts with 120 or more, that falls to about 41%.
The practical question is not whether a large cast is “wrong.” It is whether the reader can still tell whose choices are driving the scene. The underlying rank score is −0.523; it describes a pattern, not dramatic effectiveness.
How dialogue concentration changes with cast size
Within each cast-size group, what share of all dialogue belongs to its three most talkative characters?
Bars: median share · n: screenplays in each cast-size group
Choose a range, or tap or hover the chart, to inspect exact values.
View exact-value table
| Speaking characters | Scripts | Median top-three share |
|---|---|---|
| 3–19 | 73 | 71.6% |
| 20–29 | 244 | 65.1% |
| 30–39 | 354 | 62.5% |
| 40–59 | 648 | 58.2% |
| 60–79 | 350 | 52.4% |
| 80–119 | 231 | 49.3% |
| 120+ | 177 | 41.3% |
Corpus anatomy
What a middle-of-the-pack script looks like here
The center marker is the median: half the included scripts fall above it and half below. The wider bar holds the middle half of scripts. These are word and scene counts, not page counts, because MovieSum does not preserve stable pagination.
Narrative words
10th · 15,87290th · 29,019
- 25th
- 18,380
- Median
- 21,579
- 75th
- 25,179
Scenes
10th · 6890th · 211
- 25th
- 97
- Median
- 130
- 75th
- 168
Typical scene length (words)
10th · 5390th · 189
- 25th
- 70
- Median
- 94
- 75th
- 132
Speaking characters
10th · 2690th · 110
- 25th
- 36
- Median
- 50
- 75th
- 72
Story held in longest 10% of scenes
10th · 32%90th · 46%
- 25th
- 35%
- Median
- 38%
- 75th
- 42%
Scenes with no dialogue
10th · 13%90th · 47%
- 25th
- 20%
- Median
- 29%
- 75th
- 39%
The last two measures come directly from scene content. In the median script, the longest tenth of scenes carries 38% of all action-and-dialogue words, while 29% of scenes contain no tagged dialogue. They reveal concentration and silence—not whether those choices work.
What a plot summary leaves behind
A plot summary is a map, not a miniature screenplay
The median summary keeps only 3.02% of the screenplay’s counted words. That reduction removes performance, staging, rhythm, and most scene-level cause-and-effect. The numbers below help explain what summaries discard; they should not be used to judge screenplay craft.
For scale, the median screenplay contains 20,759 narrative words and its summary contains 656. Longer scripts do not reliably receive proportionally longer summaries.
Screenplay narrative100%
Wikipedia summary3.02%
Unique summary phrases absent from the screenplay
Single words30.6%
Two-word phrases72.6%
Three-word phrases95.1%
“Absent” means the exact word sequence does not appear in the screenplay. That can be useful paraphrase or an error; this test cannot tell which. The takeaway is simpler: summaries frequently describe events in language the screenplay itself never uses.
Methodology and boundaries
How we counted—and where to be careful
What counted
We kept scripts with enough material to compare reliably: at least 20 scenes, 5,000 action-and-dialogue words, three speaking characters, and 20 dialogue blocks.
How scripts line up
Each script becomes ten equal stretches of story text, from first scene to last. Every screenplay gets one vote, so a very long script cannot overpower the rest.
What not to infer
This is not a sample of every working screenplay. Word position is not page count or screen time, and a common pattern is not automatically a good choice.
Technical notes for researchers
Charts use medians and observed quartiles rather than assuming a bell curve. The downloadable data also includes deterministic 95% bootstrap intervals, rank correlations, and log-normal fit checks.
We inspected 2,200 records, included 2,077, and excluded 123. Exclusion counts can overlap when one file misses more than one threshold.
Analysis version 1.1.0 records the thresholds, random seed, and source fingerprint in the aggregate JSON. No screenplay text, summaries, IMDb identifiers, or title-level measurements are served.
Work with the measurements
Take the distributions into your own analysis
Open a table in a spreadsheet, compare its quantiles, or build another view of the same measurements. The complete JSON retains the correlations, fit diagnostics and methodology alongside these tables.
Movement across the draft
Action share, scenes per normalized tenth and cumulative speaking-cast arrival, with empirical quantiles and published median intervals.
30 measurement rows · Analysis 1.1.0
Preview measurements and column definitions
2,077 eligible screenplays. Cast-bin counts appear beside each measure; finite observation counts for individual statistics are not published.
| Measure | P10 | P25 | Median | P75 | P90 | Median interval low | Median interval high |
|---|---|---|---|---|---|---|---|
| Action word sharefraction · 0–10% | 0.4186 | 0.5257 | 0.6382 | 0.743 | 0.8262 | 0.6318 | 0.6486 |
| Scenes per normalized tenthscenes · 0–10% | 5 | 8 | 13 | 17 | 23 | 12 | 13 |
| Speaking cast introducedfraction · 0–10% | 0.1247 | 0.1613 | 0.2174 | 0.2807 | 0.3571 | 0.2121 | 0.2211 |
| Action word sharefraction · 10–20% | 0.3419 | 0.439 | 0.5573 | 0.6598 | 0.755 | 0.5501 | 0.566 |
| Scenes per normalized tenthscenes · 10–20% | 5 | 7 | 11 | 16 | 21 | 11 | 11 |
| Speaking cast introducedfraction · 10–20% | 0.2338 | 0.2885 | 0.3571 | 0.4444 | 0.5202 | 0.3511 | 0.3636 |
| Action word sharefraction · 20–30% | 0.3339 | 0.4471 | 0.5514 | 0.6657 | 0.7656 | 0.5434 | 0.5607 |
| Scenes per normalized tenthscenes · 20–30% | 5 | 8 | 11 | 16 | 22 | 11 | 12 |
| Speaking cast introducedfraction · 20–30% | 0.3333 | 0.3881 | 0.4651 | 0.5556 | 0.6414 | 0.4603 | 0.4727 |
| Action word sharefraction · 30–40% | 0.3418 | 0.4361 | 0.5474 | 0.6584 | 0.7612 | 0.5389 | 0.5569 |
| Scenes per normalized tenthscenes · 30–40% | 5 | 8 | 12 | 17 | 23 | 11 | 12 |
| Speaking cast introducedfraction · 30–40% | 0.4245 | 0.4848 | 0.56 | 0.6458 | 0.7253 | 0.5541 | 0.5667 |
| Action word sharefraction · 40–50% | 0.3345 | 0.4314 | 0.5447 | 0.6535 | 0.7569 | 0.5351 | 0.5535 |
| Scenes per normalized tenthscenes · 40–50% | 5 | 8 | 12 | 17 | 22.4 | 11 | 12 |
| Speaking cast introducedfraction · 40–50% | 0.5152 | 0.5726 | 0.65 | 0.7308 | 0.8046 | 0.6429 | 0.6552 |
| Action word sharefraction · 50–60% | 0.3396 | 0.444 | 0.5446 | 0.6617 | 0.7669 | 0.5375 | 0.5529 |
| Scenes per normalized tenthscenes · 50–60% | 5 | 8 | 12 | 17 | 23 | 12 | 12 |
| Speaking cast introducedfraction · 50–60% | 0.6026 | 0.6591 | 0.7339 | 0.8033 | 0.8681 | 0.7292 | 0.7407 |
| Action word sharefraction · 60–70% | 0.3411 | 0.4455 | 0.5572 | 0.6734 | 0.7703 | 0.5506 | 0.5654 |
| Scenes per normalized tenthscenes · 60–70% | 5 | 8 | 12 | 18 | 24 | 12 | 13 |
| Speaking cast introducedfraction · 60–70% | 0.68 | 0.7407 | 0.8056 | 0.8661 | 0.92 | 0.8 | 0.8107 |
| Action word sharefraction · 70–80% | 0.3381 | 0.4449 | 0.5629 | 0.6845 | 0.7966 | 0.5522 | 0.5755 |
| Scenes per normalized tenthscenes · 70–80% | 5 | 9 | 13 | 18 | 25 | 12 | 13 |
| Speaking cast introducedfraction · 70–80% | 0.7725 | 0.8246 | 0.8772 | 0.925 | 0.963 | 0.8718 | 0.881 |
| Action word sharefraction · 80–90% | 0.3599 | 0.4735 | 0.5987 | 0.7327 | 0.8291 | 0.5879 | 0.6102 |
| Scenes per normalized tenthscenes · 80–90% | 5 | 9 | 14 | 20 | 27 | 13 | 14 |
| Speaking cast introducedfraction · 80–90% | 0.8684 | 0.9091 | 0.9444 | 0.9737 | 1 | 0.9412 | 0.9467 |
| Action word sharefraction · 90–100% | 0.4194 | 0.5402 | 0.6608 | 0.7848 | 0.8655 | 0.6481 | 0.6696 |
| Scenes per normalized tenthscenes · 90–100% | 5 | 8 | 14 | 20 | 28 | 13 | 14 |
| Speaking cast introducedfraction · 90–100% | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
- Action word share
- Tagged action words divided by action-plus-dialogue words.
- Scenes per normalized tenth
- Scene count in a tenth, assigned by the scene's narrative-word midpoint.
- Speaking cast introduced
- Cumulative share of eventual speaking characters introduced by the end of a tenth.
Whole-screenplay distributions
Script and summary lengths, scene and cast counts, concentration, compression and novel summary phrases. Each row summarizes one measure across the corpus.
15 measurement rows · Analysis 1.1.0
Preview measurements and column definitions
2,077 eligible screenplays. Cast-bin counts appear beside each measure; finite observation counts for individual statistics are not published.
| Measure | P10 | P25 | Median | P75 | P90 | Median interval low | Median interval high |
|---|---|---|---|---|---|---|---|
| Narrative wordswords · Whole corpus | 15871.6 | 18380 | 21579 | 25179 | 29019.4 | — | — |
| Tokenized narrative wordswords · Whole corpus | 15476.4 | 17908 | 20759 | 23715 | 27087.2 | — | — |
| Scene countscenes · Whole corpus | 68 | 97 | 130 | 168 | 211 | — | — |
| Median scene lengthwords · Whole corpus | 52.5 | 70 | 94 | 132 | 189 | — | — |
| Longest tenth of scenesfraction · Whole corpus | 0.317 | 0.3455 | 0.3792 | 0.4199 | 0.4615 | — | — |
| Dialogue-free scenesfraction · Whole corpus | 0.1329 | 0.2027 | 0.2945 | 0.3852 | 0.4697 | — | — |
| Speaking-character countcharacters · Whole corpus | 26 | 36 | 50 | 72 | 110 | — | — |
| Action word sharefraction · Whole corpus | 0.3957 | 0.4859 | 0.579 | 0.6751 | 0.7398 | — | — |
| Top-three dialogue sharefraction · Whole corpus | 0.3979 | 0.4798 | 0.5718 | 0.6603 | 0.7354 | — | — |
| Scene-heading coveragefraction · Whole corpus | 0.9863 | 0.9913 | 0.9947 | 1 | 1 | — | — |
| Summary wordswords · Whole corpus | 410.6 | 538 | 656 | 714 | 801.4 | — | — |
| Summary compression ratiofraction · Whole corpus | 0.019 | 0.025 | 0.0302 | 0.0366 | 0.0443 | — | — |
| Novel summary wordsfraction · Whole corpus | 0.242 | 0.2718 | 0.3056 | 0.3426 | 0.3784 | — | — |
| Novel two-word summary phrasesfraction · Whole corpus | 0.6727 | 0.6993 | 0.726 | 0.7557 | 0.7824 | — | — |
| Novel three-word summary phrasesfraction · Whole corpus | 0.9236 | 0.9378 | 0.9511 | 0.9627 | 0.9717 | — | — |
- Narrative words
- Action-plus-dialogue word count used for screenplay progress and inclusion.
- Tokenized narrative words
- Narrative words counted by the summary-comparison tokenizer.
- Scene count
- Parsed scenes per screenplay.
- Median scene length
- Each screenplay's median action-plus-dialogue word count per scene.
- Longest tenth of scenes
- Narrative-word share in the longest 10% of scenes, rounding the scene count up.
- Dialogue-free scenes
- Share of parsed scenes with no tagged dialogue words.
- Speaking-character count
- Normalized character cues followed by at least one dialogue block per screenplay.
- Action word share
- Tagged action words divided by action-plus-dialogue words.
- Top-three dialogue share
- Dialogue-word share carried by the three busiest speaking characters.
- Scene-heading coverage
- Share of parsed scenes with an explicit scene heading.
- Summary words
- Tokenized Wikipedia summary word count per screenplay record.
- Summary compression ratio
- Summary words divided by tokenized narrative words for each screenplay.
- Novel summary words
- Share of unique summary words absent from the screenplay token sequence.
- Novel two-word summary phrases
- Share of unique summary bigrams absent from the screenplay token sequence.
- Novel three-word summary phrases
- Share of unique summary trigrams absent from the screenplay token sequence.
Cast size and dialogue concentration
Speaking-character counts and the three busiest voices' dialogue share within each published cast-size bin, retaining its script count.
14 measurement rows · Analysis 1.1.0
Preview measurements and column definitions
2,077 eligible screenplays. Cast-bin counts appear beside each measure; finite observation counts for individual statistics are not published.
| Measure | P10 | P25 | Median | P75 | P90 | Median interval low | Median interval high |
|---|---|---|---|---|---|---|---|
| Speaking-character countcharacters · 3–1973 scripts in this bin | 9.2 | 13 | 15 | 17 | 19 | — | — |
| Top-three dialogue sharefraction · 3–1973 scripts in this bin | 0.5947 | 0.6555 | 0.7159 | 0.8308 | 0.9197 | — | — |
| Speaking-character countcharacters · 20–29244 scripts in this bin | 21 | 23 | 26 | 27 | 29 | — | — |
| Top-three dialogue sharefraction · 20–29244 scripts in this bin | 0.5107 | 0.5771 | 0.6511 | 0.7365 | 0.7982 | — | — |
| Speaking-character countcharacters · 30–39354 scripts in this bin | 31 | 33 | 35 | 37 | 38.7 | — | — |
| Top-three dialogue sharefraction · 30–39354 scripts in this bin | 0.482 | 0.5346 | 0.625 | 0.6986 | 0.7554 | — | — |
| Speaking-character countcharacters · 40–59648 scripts in this bin | 41 | 44.75 | 49 | 54 | 57 | — | — |
| Top-three dialogue sharefraction · 40–59648 scripts in this bin | 0.4277 | 0.5023 | 0.582 | 0.6579 | 0.7178 | — | — |
| Speaking-character countcharacters · 60–79350 scripts in this bin | 61 | 63 | 69 | 73 | 77 | — | — |
| Top-three dialogue sharefraction · 60–79350 scripts in this bin | 0.3988 | 0.4567 | 0.5236 | 0.6016 | 0.6759 | — | — |
| Speaking-character countcharacters · 80–119231 scripts in this bin | 81 | 86 | 95 | 104 | 113 | — | — |
| Top-three dialogue sharefraction · 80–119231 scripts in this bin | 0.3543 | 0.4099 | 0.4927 | 0.5664 | 0.6284 | — | — |
| Speaking-character countcharacters · 120+177 scripts in this bin | 125 | 136 | 164 | 235 | 312.4 | — | — |
| Top-three dialogue sharefraction · 120+177 scripts in this bin | 0.2671 | 0.3427 | 0.4135 | 0.4914 | 0.5769 | — | — |
- Speaking-character count
- Normalized character cues followed by at least one dialogue block per screenplay.
- Top-three dialogue share
- Dialogue-word share carried by the three busiest speaking characters.
For local analysis, save a CSV and its table dictionary in the same folder, keeping both filenames. The complete table metadata describes all three CSVs together.
Shares are fractions from 0 to 1; multiply by 100 to display a percentage.
P25–P75 describes variation between screenplays. The 95% interval estimates uncertainty in a median; it is not a range containing 95% of scripts. Blank interval cells mean no interval was published, not zero.
Corpus scripts counts eligible screenplays, not the number of finite observations used for every statistic. Group scripts is published only for cast-size bins; per-tenth observation counts are not present in the source aggregate.
Column definitions
- metric
- Stable measure key; definitions appear in the table metadata and preview.
- unit
- Fraction, words, scenes or characters. Quantile interpolation can produce fractional counts.
- group
- Normalized word-progress range, cast-size bin, or all for a whole-corpus distribution.
- corpus_scripts
- Eligible screenplay count; not a statistic's finite observation count.
- group_scripts
- Published cast-bin screenplay count. Empty for other groups whose finite observation counts are not published.
- p10
- Empirical 10th percentile of per-screenplay measurements.
- p25
- Empirical 25th percentile of per-screenplay measurements.
- median
- Empirical 50th percentile of per-screenplay measurements.
- p75
- Empirical 75th percentile of per-screenplay measurements.
- p90
- Empirical 90th percentile of per-screenplay measurements.
- median_ci95_low
- Lower bound of the published percentile-bootstrap 95% median interval; empty when unavailable.
- median_ci95_high
- Upper bound of the published percentile-bootstrap 95% median interval; empty when unavailable.
- analysis_version
- Version of the source analysis; exporting a table does not rerun that analysis.
- source_input_sha256
- SHA-256 input fingerprint from the published aggregate.
- canonical_url
- Atlas landing page and dataset identity to retain with a reused table.
- license_url
- Published aggregate license; not a new reuse grant.
- source_credit
- Source creators, source URL and adaptation attribution carried with each exported row.
Measure the draft you actually have
Corpus patterns are context. Your screenplay is the decision.
Put one draft on these exact distributions with the pacing map—or use the screenplay analyzer for the scene, character, location, and dialogue shape of one draft, or compare two revisions without flattening the screenplay into a generic text diff.