Methodology

Analysis pipeline (9 stages)

Each analysis you request runs these 9 stages in order. Stages 1-2 are deterministic (no LLM). Stages 3-9 call Gemini with versioned, traceable prompts.

1. 1Parser. Text extraction from PDF / Word / TXT with deep cleanup: page numbers, page-number folios glued to next line's action, hyphen-broken words across line breaks, letter-spaced text, repeated headers and footers. No LLM.
2. 2Stats. Scene segmentation with heuristics calibrated against a corpus of real production-grade screenplays: INT/EXT detection, special scenes (FLASHBACK, DREAM, MONTAGE), slug-line parentheticals, character-name exclusions, ambiguous headers and continuations. Word counts, interior/exterior distribution, estimated runtime. No LLM.
3. 3Per-scene analysis with Gemini 3 Flash. Summary, emotions (Plutchik), characters in scene with objective and emotional arc, dialogues with subtext, semantic field (Greimas), props and visual/sound production elements.
4. 4Global analysis with Gemini 3.1 Pro. Central conflict (McKee / Egri), typed plot points (Field / Snyder), character arcs (Truby / Campbell), thematic clusters (Greimas / Bal), structural and rhythm patterns (Mamet), aggregated emotional curves.
5. 5Revision proposals with Gemini 3.1 Pro. Flagging of possible issues in spelling, grammar and agreement, punctuation, accents, header format, character consistency, mismatched parentheses, near repetitions, temporal and location coherence, tense. The system suggests; the writer decides what to accept.
6. 6Executive synthesis with Gemini 3.1 Pro. Premise / controlling idea (McKee), primary genre + sub-genres with confidence and rationale, cinematic comparables (3-5 titles with explicit similarity dimensions).
7. 7Diversity and representation with Gemini 3.1 Pro. Automated Bechdel test (Alison Bechdel, 1985: ≥2 named female characters, who talk to each other, about something other than a man). Cast composition by gender detected in text. Representation notes.
8. 8Structured extraction with LangExtract (Gemini 3 Flash). Characters, emotions, locations, stage directions, relationships, objects and themes with literal citation and character offsets back to the script (zero-hallucination).
9. 9Per-scene embeddings with gemini-embedding-2 (output_dim 768 via MRL truncation, auto-normalized). The RAG chat embeds every question and retrieves the top-5 most similar scenes by cosine similarity before calling Gemini 3.1 Pro — traceable answers, scoped to the analysis, or explicit "not in the analyzed script".

Models in use

Configurable via backend env vars. Allows swapping between Gemini Flash (fast, cheap) and Gemini Pro (deep reasoning) per stage.

• Per-scene analysis, structured extraction (LangExtract): gemini-3-flash-preview. Each scene is manageable context; speed and cost matter.
• Global analysis, correction, synthesis, diversity, RAG chat: gemini-3.1-pro-preview. Long context + reasoning over the whole script.
• Per-scene embeddings: gemini-embedding-2 (new google-genai SDK). 768 dimensions via MRL truncation, auto-normalized for cosine.

Derived analyses (no extra LLM)

Once the 9-stage pipeline finishes, several views compute additional metrics on the data already produced. They DO NOT consume tokens nor add per-analysis cost — they are pure computations on the resulting JSON. That makes them fast and repeatable without waiting.

• Quality score 0-100. Weighted synthesis of the severities of detected issues in the correction stage. 100 = no issues; each HIGH subtracts 4 points, MEDIUM 1.5, LOW 0.3. Not a subjective rating of the script — a quantitative measure of pending formal correction.
• Global subtext. % of dialogue lines with explicit subtext per the per-scene analysis. Categories: Subtle (>60%), Balanced (30-60%), Explicit (<30%). Useful to understand the script's register without judging good or bad.
• Filmability. Detection of introspective markers ("thinks", "remembers", "imagines", "feels"…) in action/description. Heavily penalizes internal text. Useful for novel adaptations where introspective density isn't directly visualizable.
• Production complexity. Score 0-100 calibrated on 4 axes (unique locations, special scenes, cast size, unique props), with different curves per axis. Doesn't predict budget in euros; places the project in a band (micro-budget / indie / studio / blockbuster).
• Production breakdown. List of physical elements (props, vehicles, locations) extracted from the per-scene analysis, grouped and counted. Useful as a starting point for a manual production breakdown.
• Dialogue balance. Share of total dialogue lines per character. Detects absorbing leads (>50% of dialogue) and underdeveloped characters (several scenes but few lines).
• Version diff. Compares two analyses of the same script via cosine similarity of scene embeddings. Identifies kept, moved, added, and removed scenes. Useful when iterating a script to see what actually changed.
• Character co-occurrences. Lightweight graph of which characters appear together in scenes, based on per-scene rosters from the global analysis.
• Duplicate scenes detection. Pairs of scenes with cosine similarity > 0.85 on embeddings. Detects thematic repetitions or near-identical scenes that may be asking for a merge.
• Roster coherence. Detects characters that have dialogue lines in a scene but don't appear in its declared roster. Usually signals an oversight in the analysis or an inconsistency in the script itself.

Diversity and representation

Festival window

What the system does NOT do (deliberately)

Some decisions belong exclusively to professional human judgment. The system doesn't replace them.

• Does not suggest casting or actor profiles. Casting choices depend on the director, producer, audition process, and actual talent availability. Any system that predicts "ideal actor for this character" oversimplifies a complex process and unconsciously biases toward known faces.
• Does not predict commercial success or box office. That would require marketing, distribution, and market data that don't derive from the script.
• Does not pass aesthetic judgment. The system says "this script has 5 special scenes with effects" but not "this script is good/bad". Taste belongs to the reader.
• Does not rewrite. The Tweaks tab proposes revisions, but does NOT apply automatic changes to the text. The screenwriter keeps full control.
• Does not substitute a human script advisor. That's why we offer the Expert advisor tab: to connect you with a professional when automated analysis falls short.

Bibliography

References in APA format with an explanation of what we take from each author. Citations in the results views link to their entry here.

• Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., & Dodds, P. S. (2016). The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science, 5(31).

  What we use from this author

  Estudio empírico (Project Gutenberg, 1.327 obras) que demuestra que la mayoría de relatos siguen una de 6 formas emocionales arquetípicas (rags-to-riches, tragedy, man-in-a-hole, etc.). Base para la clasificación de la curva emocional del guion.

  #reagan-2016

• Moretti, F. (2011). Network theory, plot analysis. New Left Review, 68.

  What we use from this author

  Análisis de la trama como red de interacciones entre personajes. Inspira nuestra vista de co-apariciones de personajes y la centralidad relativa en el cast.

  #moretti-2011

• Truby, J. (2007). The anatomy of story: 22 steps to becoming a master storyteller. Faber & Faber.

  What we use from this author

  Arco moral del personaje en 7 pasos: necesidad, deseo, oponente, plan, gran batalla, revelación, nuevo equilibrio. Esquema que estructura los character arcs reportados en el análisis global.

  #truby-2007

• Snyder, B. (2005). Save the cat! The last book on screenwriting you'll ever need. Michael Wiese Productions.

  What we use from this author

  Beat sheet de 15 beats (opening image, theme stated, catalyst, midpoint, all is lost, etc.). Lo usamos junto a Field para reconocer plot points específicos en el análisis estructural.

  #snyder-2005

• McKee, R. (1997). Story: Substance, structure, style, and the principles of screenwriting. ReganBooks.

  What we use from this author

  Concepto de controlling idea (premise positiva o negativa con value charge) que identifica la idea controladora del guion. También su definición de conflicto central (interno / interpersonal / extra-personal).

  #mckee-1997

• Mamet, D. (1997). True and false: Heresy and common sense for the actor. Pantheon Books.

  What we use from this author

  Principios de economía dramática: cada escena debe tener un objetivo claro y un cambio. Lo usamos para evaluar el ritmo y para flaggear escenas que no avanzan la trama.

  #mamet-1997

• Howard, D., & Mabley, E. (1993). The tools of screenwriting: A writer's guide to the craft and elements of a screenplay. St. Martin's Griffin.

  What we use from this author

  Las 14 herramientas (tema, premise, exposición, etc.). Nos da el vocabulario operacional con el que pedimos al LLM que describa la mecánica de cada escena.

  #howard-mabley-1993

• Vogler, C. (1992). The writer's journey: Mythic structure for writers. Michael Wiese Productions.

  What we use from this author

  Adaptación práctica del viaje del héroe de Campbell para guionistas. 12 etapas operacionalizadas que nos sirven para detectar patrones míticos en arcos de personaje.

  #vogler-1992

• Ekman, P. (1992). An argument for basic emotions. Cognition and Emotion.

  What we use from this author

  6 emociones básicas universales (alegría, tristeza, miedo, ira, sorpresa, asco). Marco alternativo a Plutchik que reportamos cuando se identifican emociones más fundamentales.

  #ekman-1992

• Seger, L. (1987). Making a good script great. Samuel French.

  What we use from this author

  Manual práctico de reescritura. Aporta criterios concretos para reconocer plot points débiles, escenas duplicadas funcionalmente y subtramas que no resuelven. Usado en la pestaña Ajustes.

  #seger-1987

• Bal, M. (1985). Narratology: Introduction to the theory of narrative. University of Toronto Press.

  What we use from this author

  Distinción entre fábula (eventos), historia (orden) y texto (narración). Complementa Greimas para mapear los clusters temáticos a su despliegue narrativo en el guion.

  #bal-1985

• Bechdel, A. (1985). Dykes to watch out for [comic strip: "The Rule"]. Self-published.

  What we use from this author

  Origen del test de Bechdel: una obra cumple los 3 criterios si (1) tiene al menos 2 personajes femeninos con nombre, (2) que hablan entre sí, (3) sobre algo que no sea un hombre. Test mínimo de presencia femenina, no certifica representación rica.

  #bechdel-1985

• Bordwell, D. (1985). Narration in the fiction film. University of Wisconsin Press.

  What we use from this author

  Modelo de narración clásica de Hollywood: la historia se construye con una cadena de causa-efecto donde cada escena motiva la siguiente por la acción/decisión de un personaje. Lo usamos como marco teórico para evaluar la cadena causal escena-a-escena (enlaces fuertes vs. débiles vs. rotos).

  #bordwell-1985

• Plutchik, R. (1980). A general psychoevolutionary theory of emotion. Academic Press.

  What we use from this author

  Rueda de las 8 emociones primarias (alegría, confianza, miedo, sorpresa, tristeza, asco, ira, anticipación) con intensidades y combinaciones. Es la taxonomía que pedimos al LLM para etiquetar emociones por escena.

  #plutchik-1980

• Field, S. (1979). Screenplay: The foundations of screenwriting. Dell.

  What we use from this author

  Estructura clásica en 3 actos con plot points en aprox. página 25 (Plot Point 1) y página 85 (Plot Point 2) en un largometraje estándar. Lo usamos para tipar los plot points detectados.

  #field-1979

• Genette, G. (1972). Figures III (incluye "Discours du récit"). Seuil.

  What we use from this author

  Tipología de tiempo narrativo: orden (anacronías), duración (escena, sumario, pausa, elipsis) y frecuencia. Lo usamos para analizar el ritmo del guion a partir de la longitud de escenas y los plot points.

  #genette-1972

• Greimas, A. J. (1966). Sémantique structurale. Larousse.

  What we use from this author

  Modelo actancial (sujeto, objeto, destinador, destinatario, ayudante, oponente) y campo semántico como agrupación de oposiciones binarias (vida/muerte, naturaleza/cultura). Lo usamos para identificar los clusters temáticos del guion.

  #greimas-1966

• Jung, C. G. (1959). The archetypes and the collective unconscious. Pantheon Books.

  What we use from this author

  Arquetipos del inconsciente colectivo (héroe, sombra, anima/animus, sabio, etc.). Vocabulario que el LLM utiliza al describir personajes prototípicos cuando se ajustan al esquema.

  #jung-1959

• Campbell, J. (1949). The hero with a thousand faces. Pantheon Books.

  What we use from this author

  Monomito: estructura universal del viaje del héroe (separación, iniciación, retorno). Base teórica de los arcos de personaje cuando detectamos viaje transformador.

  #campbell-1949

• Egri, L. (1942). The art of dramatic writing. Simon & Schuster.

  What we use from this author

  Premise dramática como motor del conflicto, y tridimensionalidad del personaje (fisiología, sociología, psicología). Lo usamos junto a McKee para evaluar el conflicto central.

  #egri-1942