LLMs can draft “stories” at the press of a button. They sound like fantasy, science fiction, literary realism—whatever you ask for. But when you look at them as stories rather than as text, the cracks show up fast:
- Arcs flatten.
- Turning points blur.
- Emotions skate along the surface.
- The theme you thought you were asking for quietly drifts.
A recent EMNLP 2025 survey, “A Survey on LLMs for Story Generation,” inadvertently documents this problem in detail. It reviews the latest LLM-based story systems—independent generators, author-assistance tools, educational and therapeutic story apps—and still finds persistent gaps in narrative coherence, emotional depth, and evaluation.
What these systems share is a focus on text intelligence—getting the words to look right—without an explicit model of narrative intelligence.
The Dramatica platform (Dramatica theory → Subtxt → Narrova → NCP) was built to fill exactly that gap.
What “Text Intelligence” Keeps Missing
The EMNLP survey organizes almost all LLM story work into two buckets:
- Independent story generation – the model is treated as the primary author.
- Author assistance – the model acts as co-author, coach, prototype machine, or brainstorm partner.
Across both, you see the same technical patterns:
- Prompt templates and complex prompt stacks.
- Outline-to-draft flows that try to give the model a skeleton to follow.
- Multi-agent setups where different models act as critics, planners, or action selectors.
- UX-heavy tools for character chat, educational storytelling, or game prototyping.
Despite all that engineering effort, the survey’s own findings highlight the same recurring gaps:
-
No agreed benchmarks for narrative quality.
Evaluation is fragmented: small user studies, task-specific anecdotes, and one-off experiments. There is no shared notion of what “good story” means structurally. -
Metrics that measure similarity, not story.
BLEU, ROUGE, BERTScore, and generic LLM-as-a-judge ratings tell you whether output is similar or superficially “good,” but not whether it has a coherent argument, balanced perspectives, or a meaningful arc. -
Fragile arcs, weak turning points, thematic drift.
Even when systems scaffold LLMs with outlines, emotion tags, or knowledge graphs, they still struggle to maintain strong story arcs, clear turning points, and consistent themes across longer works.
In short, current systems optimize surface text without a durable representation of narrative intent. They are trying to fix story problems using tools that only understand words.
Narrative Intelligence, the Dramatica Way
Text intelligence asks:
“What’s the next plausible token?”
Narrative intelligence asks:
-
What inequity sits at the heart of this Story?
What’s out of balance in the world of the story—internally, externally, or both—and why does it matter? -
Which Perspectives are exploring that inequity?
How do the Main Character, Influence Character, Objective Story, and Relationship Story each view the same underlying problem? -
How do those Perspectives evolve?
How do Signposts and Journeys trace the transformation (or non-transformation) of those perspectives over time? Where do the true turning points happen? -
What argument is the Story actually making about how to resolve the inequity?
Given the chosen Problem/Solution pair and the story’s Dynamics (change vs steadfast, success vs failure, etc.), what conclusion are we asking the audience to draw?
Dramatica encodes these answers as a Storyform—a precise configuration of:
- Dynamics – global choices like change/steadfast, success/failure, good/bad, story driver, and more.
- Storypoints – Domains, Concerns, Issues, Problems, Solutions, and related items organized in quad structures.
- Storybeats – Signposts and Journeys that specify how conflict progresses through each throughline.
This Storyform is a model of narrative intelligence. It is how the Dramatica platform knows what must happen structurally before a single sentence of text is written.
The Platform: Dramatica → Subtxt → Narrova → NCP
The Dramatica platform turns that narrative model into a working ecosystem:
- Dramatica theory – the Story Mind model and the Storyform itself.
- Subtxt – where authors work in meaning, not just words; Storyforming decisions cascade down into Storybeats and scene-level guidance.
- Narrova – a multi-agent narrative system aligned with Dramatica’s communication stages: Storyforming, Story Encoding, Storyweaving, Story Reception.
- NCP (Narrative Context Protocol) – an open standard that makes Storyforms portable, so tools and agents can share the same narrative intent.
Together, they do what generic LLM stacks cannot: reason about, generate, and evaluate stories as stories, not just as streams of tokens.
Let’s look at those layers more closely.
Subtxt: Writing in Storybeats
Subtxt is where you stop wrestling with prompts and start authoring a Storyform-driven narrative.
Capture the inequity, Domains, and Storypoints that define the Storyform.
In Subtxt you begin by articulating the core inequity—what is fundamentally out of balance. You map that inequity into Dramatica’s four Domains (Situation, Activity, Fixed Attitude, Manipulation) and assign Storypoints (Concerns, Issues, Problem/Solution, etc.) for each throughline. Instead of a hand-wavy premise (“a cop questions his loyalty”), you get a precise Storyform that encodes what this story is about at a structural level.
See Signposts and Journeys as constraints that keep scenes honest.
Once the Storyform is set, Subtxt lays out the Signposts and Journeys that must unfold in each throughline. These are not mere section titles; they are constraints on what each sequence of scenes has to accomplish in terms of conflict and progression. When you draft a scene, you are working inside a specific Storybeat, with a clear sense of what kind of change must occur for the overall argument to remain coherent.
Invoke AI helpers that stay within the Storyform so drafts serve the intended argument.
Subtxt’s AI assistance doesn’t just “continue the text.” It generates material that is conditioned on the Storyform and the active Storybeat. When you ask for a scene or line of dialogue, the system knows which throughline is in play, what type of conflict belongs there, and what cannot happen without breaking the structure. The result is AI-generated text that supports your narrative argument instead of drifting away from it—the exact failure mode many surveyed systems struggle with.
Narrova: Agents by Narrative Function
Narrova is the multi-agent narrative intelligence layer, but it doesn’t define agents by ad hoc roles like “writer,” “critic,” or “action picker.” Instead, it aligns them with Dramatica’s four stages of communication.
Storyforming agents steward the Storyform.
These agents help you create and maintain the Storyform itself: clarifying the inequity, choosing Dynamics, and aligning Domains and Storypoints. When you change a high-level intent—say, flipping from success to failure—Storyforming agents propagate that change through the structure, keeping everything legal and coherent.
Story Encoding agents turn structure into concrete material.
Encoding agents map abstract structural elements into the fictional world: characters, settings, incidents. They help answer questions like, “Who embodies this Issue?” or “What specific situation best expresses this Problem/Solution pair?” They transform Storypoints into things an audience can see, hear, and feel.
Storyweaving agents manage revelation and timing.
Weaving agents decide how and when to reveal information. They orchestrate cut points between throughlines, control pacing, and ensure that the order in which the audience receives information serves the intended emotional and thematic progression. Even a perfect structure can fall flat if it’s revealed in the wrong sequence; Storyweaving agents exist to prevent that.
Story Reception agents reason about audience processing.
Reception agents take the audience’s perspective, asking: “Given what we’ve shown, what does the audience believe? What are they expecting? Is the argument landing?” They help you evaluate whether the intended meaning is likely to be the received meaning—and where you might need to adjust structure, encoding, or weaving to align them.
Because every agent shares the same Storyform via NCP, collaboration stays aligned.
All Narrova agents read and write the same Storyform, encoded in NCP. When Storyforming changes a structural choice, Story Encoding and Storyweaving immediately operate on the updated model. This stands in stark contrast to typical multi-LLM setups in the survey, where each model has a partial or local view and coherence is patched after the fact.
NCP: Portable Narrative Intent
The Narrative Context Protocol (NCP) turns Dramatica’s Storyform into a portable, machine-readable asset.
Any system can read an NCP document to understand intended structure.
An NCP document encodes the Storyform: inequity, Perspectives, Storypoints, Storybeats, and Dynamics. Any compatible tool—writing environment, game engine, visualization system, evaluator—can load it and instantly understand the intended narrative structure, independent of any particular telling.
That structure can guide generation, evaluation, adaptation, or visualization.
With NCP in place, structure becomes a shared control layer:
- A text generator can enforce Storyform constraints while drafting scenes.
- An evaluation pipeline can compare a draft against the encoded Storypoints to identify gaps or contradictions.
- An interactive narrative system can adapt player choices while staying within the intended story argument.
- A visualization or analytics tool can map beats to scenes, locations, timelines, or emotional graphs.
Story intent moves consistently across tools—writing apps, game engines, analysis pipelines.
Instead of re-encoding your story logic for each new tool, you carry a single NCP file. Subtxt, Narrova, and any NCP-aware system can all operate on that same narrative intent. In a research landscape full of bespoke formats and one-off datasets, NCP offers a stable backbone—a way to give many agents and tools access to the same story brain.
Stepping Beyond Text Intelligence
The EMNLP survey shows a field that is clearly moving in Dramatica’s direction, even if it doesn’t name it that way:
- There is a call for better metrics and benchmarks that actually measure story, not just similarity.
- There is recognition that LLMs struggle with story arcs, turning points, and affective trajectories, even with scaffolding.
- There are early moves toward planning, multi-agent systems, and constraints—all attempts to graft structure onto text-first models.
The Dramatica platform already does what these systems are reaching for:
- It starts from a formal narrative model (the Storyform).
- It gives authors a deep workspace in Subtxt to work in meaning and structure.
- It coordinates multi-agent narrative work via Narrova, with roles defined by narrative function.
- It makes narrative intent portable and interoperable through NCP.
If you are serious about story and AI, the pattern is hard to ignore:
- Prompt engineering alone will not fix narrative problems.
- Multi-agent orchestration without a shared story brain just multiplies the confusion.
- Evaluation without a theory of “good story” will always chase its own tail.
The alternative is straightforward:
Stop treating story as text to be completed.
Start treating it as narrative to be modeled.
And that path, sooner or later, leads to Dramatica.
Read the Full Whitepaper
This blog post is a condensed view. The full whitepaper:
- Gives a structured comparison between the EMNLP survey’s taxonomy and the Dramatica platform.
- Details structural metrics and benchmarking strategies using NCP.
- Explains how Narrova’s agent design differs fundamentally from current multi-LLM architectures.
- Lays out a vision for narrative intelligence as the foundation for serious story AI.
You can download it here: