Beyond Monolithic Prompts
Inside Fabula’s Compositional AI for Narrative Analysis
Standard Large Language Model (LLM) prompting, no matter how detailed, consistently fails at complex narrative analysis—yielding generic summaries instead of deep structural and psychological insight. This failure highlights a fundamental disconnect: while traditional broadcaster websites model the schedule, fan communities and audiences model the story—the intricate web of characters, storylines, and places they actually care about. This gap reveals an unmet strategic need for an industrial-scale solution that can programmatically create the rich, interconnected story models audiences demand.
For Fabula, “Extraction Intelligence” is engineered to solve this problem. Think of it like a showrunner for a TV series, coordinating specialized virtual experts—a director for performance, a writer for plot, an editor for pacing—to deconstruct and represent a narrative with human-level depth. Advanced narrative analysis is not a prompting problem. It is an architectural one, requiring a purpose-built compositional cognitive system.
Why Generic Prompts Produce Shallow Analysis
Using a single, monolithic prompt for complex analysis is like asking a general practitioner to perform brain surgery. The LLM, lacking a specialized role or analytical framework, defaults to summarizing what’s explicit in the text—completely missing the subtext, psychology, and structural significance of the narrative. It lists what happens but fails to explain why it matters.
This difference in depth is not subtle. A generic prompt might describe a character by their job, while Fabula’s system analyzes their behavior, contradictions, and motivations. The contrast is stark:
Generic Prompt Output
”Josh is a man who works in an office.”
Fabula’s Persona-Driven Output
“Josh paces the Roosevelt Room with barely controlled energy, hands gesturing emphatically even as his voice maintains calculated calm. His rapid-fire dialogue masks anxiety—words become armor against admitting uncertainty. The performance is for himself as much as others: project confidence until it becomes real.”
The generic output is factually correct but analytically useless. It leads to inconsistent, shallow analysis that fails to build coherent understanding. It sees the words on the page but misses the story being told.
Compositional Cognitive Architecture
A “compositional cognitive architecture” is a modular system combining specialized LLM roles, each with a distinct purpose, to achieve analytical depth that a single role cannot. Our approach stands as a robust engineering solution to the complex challenges of formalizing narrative functions, conflict, and character roles—problems that academics in computational narratology have theorized for years with ontologies like Drammar and ProppOnto.
Instead of one giant prompt, Fabula uses an assembly of precisely defined personas and templates that work in concert. This system has three core components:
Base Personas: Define the core analytical stance of the LLM for a given task
Specialist Modifiers: Layer on orthogonal domain expertise
Context and Style Templates: Enforce evidence grounding and output quality
This modular approach doesn’t just produce better results; it creates a system that is fundamentally more maintainable, extensible, and predictable. Each component can be versioned, tested, and optimized independently.
Personas in Action
Fabula’s architecture functions like a hand-picked team of specialized analysts. By combining these virtual experts, the system deconstructs narrative with rigor and insight impossible with a generic approach.
3.1. The Foundation: Base Personas
Base Personas define the fundamental analytical approach. Here’s how we engineer the Master_narrative_analyst:
You are a master narrative analyst with encyclopedic knowledge
of story structure, character development, and dramatic theory.
Core Methodology:
• Structural Recognition: Instantly identify acts, beats,
turning points, and pacing patterns
• Character Psychology: Deep understanding of motivation,
conflict, and transformation arcs
• Thematic Integration: Connect surface events to deeper
meaning and universal themes
• Evidence-Based Analysis: Ground all interpretations in
specific textual evidence
Your Analytical Standards:
- Every element must earn its place through narrative significance
- Character actions reveal psychology; plot events reveal theme
- Ambiguity is acknowledged but reasoned interpretations are provided
This persona “sees the forest AND the trees”—applying established narrative frameworks to understand holistic structure while connecting disparate elements into a unified whole.
For complex decisions requiring explicit reasoning, we deploy the Chain_of_thought_reasoner with a mandatory five-step process:
Mandatory Process:
1. Survey: Take in all available information and identify
key analytical challenges
2. Plan: Outline your analytical approach and decision-making criteria
3. Execute: Work through the analysis systematically,
showing your reasoning
4. Validate: Check conclusions against evidence and identify
potential weaknesses
5. Synthesize: Present final structured output with
confidence levels
Your Analytical Standards:
- Transparency in reasoning process
- Acknowledgment of limitations and uncertainties
- Systematic consideration of alternatives
This structured approach transforms entity resolution—determining whether “The Doctor” in episode 3 is the same entity as “Doctor” in episode 7—from guesswork into systematic analysis with visible logic.
Expert Modifiers for Deeper Insight
Specialist Modifiers layer on top of base personas to add domain-specific expertise. Each modifier is a focused instruction set that reshapes how the base persona approaches its task.
The “Performance Analyst”: Reading Subtext and Psychological Depth
The Performance_analyst_modifier transforms our system into a virtual acting coach:
PERFORMANCE ANALYST SPECIALIZATION:
You read between the lines of dialogue and action to understand
the full dramatic performance. You excel at:
- Interpreting subtext, body language, and unspoken communication
- Identifying moments where action contradicts dialogue
(revealing deeper truth)
- Understanding character masks, deception, and hidden motivations
- Analyzing power dynamics and relationship tensions
- Recognizing performance layers and emotional complexity
Apply this lens: “What is really happening beneath the surface
of what’s explicitly stated?”
This modifier is crucial for enriching character profiles. When analyzing a scene where a character says “I’m fine” while their hands shake, a generic LLM sees dialogue. The Performance_analyst_modifier sees contradiction—the gap between performance and reality that reveals psychological truth.
The “Story Surgeon”: Applying the Significance Filter
The Story_surgeon_modifier brings surgical precision to narrative analysis:
STORY SURGEON SPECIALIZATION:
You bring surgical precision to narrative analysis. Every element
must justify its existence through plot advancement, character
development, or thematic resonance. You excel at:
- Identifying redundant or undermotivated elements
- Diagnosing structural problems and pacing issues
- Distinguishing between essential story beats and filler
- Optimizing narrative efficiency without losing meaning
Apply this lens: “Does this element serve the story’s core mission,
or is it narrative fat?”
This directly addresses a fundamental problem in computational drama: how to handle the narrative requirement that characters move between locations without forcing systems to process every boring step. The Story_surgeon identifies and merges trivial transitions—”Josh walks to his office”—while preserving causally significant moments, maintaining a high signal-to-noise ratio in the event graph.
3.3. Enforcing Quality: Style Templates
Specialist knowledge means nothing if the output is bland corporate prose. Fabula’s Style_propulsive_prose template enforces dramatic energy:
PROSE STYLE REQUIREMENTS:
Write with immediacy and dramatic energy. Absorb the emotional
register of the source material and reflect it in your analytical voice.
CORE PRINCIPLES:
• Propulsive Verbs: Lead with strong action words that drive
forward momentum
• Immediate Reality: Ground descriptions in the specific
dramatic moment
• Active Construction: Subject acts on object; avoid passive distancing
• Concrete Specificity: Choose precise details over vague generalizations
FORBIDDEN PHRASES:
❌ “In this scene...” / “This part of the narrative...”
❌ “We observe that...” / “It can be seen that...”
❌ “The text reveals...” / “The narrative shows...”
TRANSFORM WEAK → STRONG:
Instead of: “In this scene, Josh appears frustrated”
Write: “Josh’s jaw tightens, frustration bleeding through
forced diplomacy”
Instead of: “The dialogue reveals character conflict”
Write: “Words become weapons as old alliances fracture”
This isn’t stylistic flourish—it’s precision engineering. By banning distancing phrases and enforcing active construction, we force the LLM to engage with the dramatic reality of the scene rather than hide behind analytical abstraction.
Beyond Plot: Achieving Psychological Inference
The true power of this architecture emerges when analyzing character psychology. Our data model tracks granular, event-level psychological states:
class AgentParticipationV2:
agent_uuid: str
event_uuid: str
observed_status: str # 20+ word performance description
emotional_state_at_event: str
goals_at_event: List[str]
beliefs_at_event: List[str]This is our practical implementation of the Belief-Desire-Intention (BDI) model, a well-established framework for modeling agent behavior based on Michael Bratman’s theory of practical reasoning. The Performance_analyst_modifier populates these fields by inferring what cannot be directly observed—the inner world driving external behavior.
Consider this transformation:
Script Text:
INT. OVAL OFFICE - NIGHT
BARTLET stands alone, staring at Jackson’s portrait.
His jaw is set.
BARTLET: Get me the Joint Chiefs.Generic Analysis:
“Bartlet orders military action”Fabula’s Psychological Extraction:
{
“observed_status”: “Bartlet stands alone before Jackson’s portrait,
jaw set in resolution that borders on rigidity. His voice is
clipped, brooking no debate. The hour—3 AM—and his physical
isolation emphasize the loneliness of command.”,
“emotional_state_at_event”: “Grim resolve layered over moral anguish”,
“goals_at_event”: [
“Take decisive military action before the window closes”,
“Project unshakeable confidence to Leo despite internal doubt”
],
“beliefs_at_event”: [
“Presidential authority requires the appearance of certainty”,
“This decision is morally justified despite personal cost”,
“History will judge this moment (hence Jackson’s portrait)”
]
}Notice what Fabula doesn’t do: it doesn’t just extract “Bartlet” and “orders” and “military.” It performs dramatic analysis—reading subtext, inferring beliefs from symbolic details (Jackson’s portrait), understanding that goals can be in tension with each other.
This depth is enforced through strict quality controls: descriptions of psychological state must be 20+ words to prevent generic labels like “angry” or “determined.” The system is forced to engage with specificity.
Maintainability at Scale
This architecture solves the fundamental engineering problem of AI narrative analysis: how to achieve both depth and reliability at scale.
Modularity enables independent optimization:
Update the Story_surgeon’s pruning logic without touching character analysis
Enhance the Performance_analyst’s subtext detection without breaking continuity tracking
Version control each component independently
Specialization enables targeted quality assurance:
Test the Chain_of_thought_reasoner’s entity resolution against ground truth datasets
Validate the Performance_analyst’s emotional state extraction through expert review
Measure the Story_surgeon’s signal-to-noise improvements quantitatively
Composition enables emergent capability:
Master_narrative_analyst + Performance_analyst_modifier + Style_propulsive_prose = psychological character profiles
Systematic_extractor + Technical_precision_modifier = reliable schema-compliant extraction
Creative_interpreter + Cultural_archaeologist_modifier + Genre_theorist_modifier = thematically rich story bibles
The system becomes greater than the sum of its parts—not through prompt alchemy, but through rigorous engineering of cognitive division of labor.
A Maintainable and Extensible Intelligence
Fabula’s success in achieving deep narrative understanding stems from a foundational shift in approach: treating narrative analysis as a specialized cognitive task that requires a team of virtual experts, not a single, all-purpose prompt.
The compositional architecture of Base Personas, Specialist Modifiers, and Style Templates creates a system where each function has a clear identity and purpose. This modularity results in an AI system that is not only more powerful but also more maintainable, extensible, and capable of producing consistently deep narrative insights.
By moving beyond monolithic prompts, we can build AI that doesn’t just process stories, but truly begins to understand them—one precisely engineered persona at a time.