--- title: Story Prototype in Narrative AI url: https://www.emergentmind.com/topics/story-prototype type: topic --- # Story Prototype in Narrative AI Searching arXiv for the cited and closely related papers on story prototypes and narrative generation. arXiv search query: "Story Prototype narrative generation knowledge graph story visualization consistency image-grounded story generation procedural emotional arc premise synthesis" A story prototype is a compact but operative narrative specification used to organize story logic before, beneath, or alongside full realization in prose, images, gameplay, or collaborative interaction. In recent arXiv work, the term spans several distinct but related forms: a genre-agnostic internal representation of characters, events, and scenes as a persistent knowledge graph; a modular premise design assembled from reusable narrative components; a stage-structured genre pattern for interactive composition; a plan iteratively revised by multiple critics; and a medium-fidelity or game-facing artifact that tests narrative interaction directly in use [2509.26461][2406.05690][2405.06685][2410.02428][2408.13962][2508.02132]. Across these variants, the common function is to separate core narrative organization from its eventual stylistic, visual, or performative realization. ## 1. Conceptual scope and major forms Recent usage does not treat a story prototype as a single formalism. Rather, it denotes a family of intermediate narrative objects that constrain, scaffold, or materialize story structure at different levels of abstraction. | Research line | Prototype form | Primary function | |---|---|---| | CreAgentive | Dual knowledge graph with semantic triples | Decouple story logic from genre/style realization | | MoPS | Modular premise tuple and nested dictionary | Synthesize diverse, controllable premises | | PatternTeller | Ordered genre-stage pattern | Enforce thematic consistency | | CritiCS | Refinable story plan and later text draft | Increase creativity and engagement while preserving coherence | | Neurodiverse picture-book system | Medium-fidelity hybrid digital/physical prototype | Support collaborative co-creation and social interaction | | Emotional-arc ARPG | Rise/Fall-labeled branching DAG | Align story progression with gameplay dynamics | The strongest unifying claim in this literature is negative: a story prototype is not reducible to a finished draft. CreAgentive explicitly defines its Story Prototype as “not a textual outline, script, or draft” but as a “text-independent abstraction layer” that captures “the story itself” through characters, causal chains, world constraints, and temporal evolution [2509.26461]. Other systems retain more surface form, but still keep the prototype distinct from final output: MoPS separates a premise design from the sentence that verbalizes it, PatternTeller separates a genre pattern from the scenes generated under it, and CritiCS separates a plan from the eventual long-form story [2406.05690][2405.06685][2410.02428]. A recurring misconception is that a prototype must already support maximal openness or branching. Several systems instead use deliberate constraint. The neurodiverse picture-book prototype encourages a linear picture-book structure, and CreAgentive currently demonstrates linear long-form generation with retrospection and foreshadowing rather than full interactive branching [2408.13962][2509.26461]. This suggests that, in current research practice, “prototype” often refers less to unrestricted narrative space than to an explicit control layer over a manageable narrative space. ## 2. Internal representations and reasoning substrates The most formalized recent use of the term appears in CreAgentive, where the Story Prototype is the core internal narrative representation of the system [2509.26461]. It is implemented as a “dual-knowledge-graph synergy structure” consisting of a Role Graph and a Plot Graph. The Role Graph is character-centric: nodes are roles with static attributes such as name, gender, and occupation, while edges encode relations such as kinship, romance, allies, and enemies, with strength, direction, and chapter labels. The Plot Graph is event- and scene-centric: events and scenes are nodes, and relations include `IN_EVENT` between roles and events and `OCCURRED_IN` between events and scenes. The prototype is maintained as chapter-level snapshots \(P^{(c)}\), enabling versioned narrative state, rollback, flashback retrieval, and synchronization across long horizons [2509.26461]. This representation is explicitly designed to solve four problems of long-form creative generation: genre-agnostic representation, decoupling story logic from surface realization, maintaining long-range coherence, and supporting complex structures such as retrospection and foreshadowing [2509.26461]. The Writing Stage does not generate the prototype; it consumes it. Recall Agent queries prior chapter prototypes for relevant events and emotional memories, Thread Agent analyzes “preset key plots and foreshadowing from subsequent chapters,” and Writer Agent transforms prototype content into genre-specific text. The same narrative logic can therefore be rendered as a novel, screenplay, or other textual form without rebuilding the underlying story structure [2509.26461]. A different but related formal lineage appears in work on story comprehension. “Non-Monotonic Reasoning and Story Comprehension” models narrative understanding through a story representation \(SR = \langle W, N, \succ \rangle\), where \(W\) is world knowledge, \(N\) is narrative observation, and \(\succ\) is a priority relation over defaults [1407.3832]. The framework combines Reasoning about Actions and Change, default reasoning, and argumentation. It defines unit arguments of property, precondition, causal, and persistence forms, constructs grounded interpretations from narrative observations, and identifies acceptable arguments as comprehension models. In this setting, the prototype is not a plan for generation but a structured substrate for elaboration and revision as new story blocks arrive [1407.3832]. These two lines differ in purpose—generation versus comprehension—but converge on a shared principle: story structure is operationally useful when represented as an explicit, revisable state rather than as raw text. A plausible implication is that story prototypes are emerging as a middle layer between symbolic narrative logic and large-model realization, regardless of whether the task is producing stories or understanding them. ## 3. Premise, pattern, and plan as prototype scaffolds Not all story prototypes are graph-structured. Several systems work with lighter but still explicit scaffolds that sit above full text and below abstract genre theory. MoPS treats the story premise itself as a modular prototype [2406.05690]. It decomposes a premise into modules \(M = \{\alpha, \beta, \gamma, \delta, \omega, \sigma\}\), where \(\alpha\) is theme, \(\beta\) background, \(\gamma\) persona, \(\delta\) event, \(\omega\) ending, and \(\sigma\) twist, so that a premise is \(p = (\alpha, \beta, \gamma, \delta, \omega, \sigma)\). Candidate values are generated conditionally in the sequence \(\alpha \rightarrow \beta \rightarrow \gamma \rightarrow \delta \rightarrow \omega \rightarrow \sigma\), stored in a nested dictionary, and traversed as key paths that function as premise designs. An LLM then integrates the selected path into a single coherent premise sentence, and a self-verification step filters obvious inconsistencies or factual errors. The framework reports 7.6k generated premises and 1k extended stories, and argues that modular synthesis yields stronger diversity, fascination, completeness, and originality than premises induced directly from LLMs or sampled from public datasets [2406.05690]. PatternTeller operates one level above the premise by using explicit genre patterns as prototypes for interactive composition [2405.06685]. These patterns are ordered lists of stage names and descriptions, induced through a two-phase process: retrieving representative examples for five “fundamental genres” and then applying most specific generalization to extract shared macro-structure. The resulting prototypes specify characteristic stage sequences for comedy, romance in the epic sense, tragedy, satire, and mystery. During generation, the user provides a premise and selects a pattern; the system expands the story stage by stage so that each generated segment fulfills the current stage description while maintaining consistency with previous stages [2405.06685]. The prototype here is neither a single sentence nor a graph, but a reusable, genre-level script for thematic progression. CritiCS makes the plan itself the prototype under active revision [2410.02428]. Its CRPLAN stage begins with a premise-derived outline and subjects that outline to multiple rounds of critique by specialized critics and a leader model. Critics target “Originality,” “Unusual story structure,” and “Unusual ending,” often under explicit personas such as psychologist, futurist, or social worker; the leader selects one critique per round, and an Evaluator later chooses the best version from the history of revisions. The refined plan then conditions CRTEXT, where two critics focused on Image and Voice revise local sentences for expressiveness [2410.02428]. In this formulation, the prototype is not static. It is an iteratively revised object whose purpose is to mediate between high-level invention and long-form realization. Taken together, these systems define a continuum of scaffolding. MoPS prototypes the seed, PatternTeller prototypes the macro-sequence, and CritiCS prototypes the outline under revision. Each treats narrative control as something that must be made explicit before full generation. ## 4. Multimodal and visual realizations Story prototypes also appear in systems where the final artifact is not only text. In visual story generation, the prototype is often distributed across latent state, memory, and prompt structure rather than stored as a single symbolic object. “Make-A-Story: Visual Memory Conditioned Consistent Story Generation” addresses story visualization as the generation of a sequence of images, one frame per sentence, from a multi-sentence textual story [2211.13319]. The core problem is maintaining actor and background consistency under reference and co-reference. The model uses an autoregressive latent diffusion architecture with a visual memory module. When generating frame \(m\), the system conditions on the current sentence embedding \(f(\mathbf{S}^m)\), previous sentence embeddings, and previous frame latents \(\mathbf{Z}^{