Dramatron: Hierarchical Script Generation

• Dramatron is a hierarchical, prompt-chaining system that leverages a 70B-parameter LLM to generate long-form creative texts such as screenplays and theatre scripts.
• It structures the writing process into discrete stages—title, character details, scene beats, location descriptions, and dialogue—to manage context effectively.
• Evaluations with professional writers highlighted its collaborative utility, iterative refinement process, and raised ethical considerations like bias and narrative coherence.

Dramatron is a hierarchical, prompt-chaining system built on a LLM (Chinchilla, 70B parameters) for the generation of long-form creative texts such as screenplays and theatre scripts. Designed explicitly for the constraints and affordances of LLMs, Dramatron structures the writing process into discrete, interdependent stages—title, character descriptions, scene beats, location details, and scene-level dialogue—capitalizing on prompt engineering rather than fine-tuning or architectural modification. It serves as a co-creative tool supporting professional writers in the ideation, drafting, and iterative refinement of scripts, while addressing context-length limitations and fostering a participatory, human-in-the-loop workflow (Mirowski et al., 2022).

1. System Architecture and Prompt-Chaining Methodology

Dramatron’s core engine leverages a strict hierarchy of abstraction to circumvent the context window constraints typical of LLMs. Rather than attempting to generate entire scripts in a single run, it structures script generation into five sequential stages:

1. Title Generation: From an initial user-provided log line $L$, Dramatron generates a script title using a task-specific prompt prefix $P_1$.
2. Character List and Descriptions: Using the log line and the generated title, a new prompt ($P_2$) elicits a set of characters and their attributes.
3. Plot Outline (Scene Beats): The model sequentially generates scene summaries, each annotated with a location and narrative arc (e.g., Exposition, Inciting Incident, Rising Action), via $P_3$.
4. Location Descriptions: For each unique location identified in the beats, Dramatron synthesizes detailed descriptions with $P_4$ for local color and consistency.
5. Scene Dialogue: Given all prior context and using previous scene information, the system generates fleshed-out, speaker-labeled dialogue for each beat with $P_5$.

Formally, Dramatron maintains a context $C_k = C_{k-1} \cup \{X_k\}$ at each stage $k$, generating new stage output $X_k$ via the LLM $M$: $X_k \sim M\left(P_k\left(C_{k-1}\right)\right)$ Each $P_k$ encodes “few-shot” prompt engineering with 1–4 representative exemplars, steering output format and stylistic features. At every stage, only the structural summaries (e.g., log line, title, brief scene table) are forwarded, maintaining a compact prompt under 2,000 tokens and avoiding memory overhead or irrelevant context dilution. Final script assembly is the concatenation of dialogue outputs across all scenes.

Supplementary safeguards detect pathological output loops, specifically in dialogue generation, by dividing output into blocks, counting repetitions, and resampling if block frequency exceeds a threshold.

2. LLM Parameters and Sampling Techniques

Dramatron employs the Chinchilla LLM (70B parameters, 1.4T token pretraining), solely controlled via prompt engineering without additional fine-tuning. At each generative stage, it uses nucleus (top-$p$) sampling with $p=0.9$ and temperature 1.0, producing up to 511 tokens per prompt. Changing the random seed at each sampling step offers alternative suggestions, supporting ideation, multi-sample selection, and collaborative exploration.

No additional scoring metric is used at runtime, aside from repetition detection. The human writer acts as the primary discriminator, selectively accepting, editing, or rejecting outputs to preserve logical consistency, coherence, and voice.

3. Interactive Workflow and User Interface

Dramatron is delivered as a Google Colab notebook, providing a streamlined, text-based interface for professional writers. At each hierarchical stage, users interact with three principal controls:

• Regenerate Suggestion: Reruns the same prompt with a new seed for alternative candidates.
• Continue Generation: Extends the last output, useful for exceeding token limits.
• Edit Manually: Allows direct modification of any field (log line, character list, scene, dialogue) in the current context.

Writers can traverse up and down the script’s abstraction hierarchy, revising upstream stages (e.g., editing the log line or character list) and cascading changes downstream as needed. Genre templating is supported by swapping or mixing exemplar “few-shot” samples in prompt prefixes, letting writers tune outputs for style, tone, and structure across formats (e.g., Greek tragedy, sci-fi).

This flexible workflow scaffolds an iterative co-creative process: log line $\rightarrow$ title $\rightarrow$ characters $\rightarrow$ scene beats $\rightarrow$ locations $\rightarrow$ dialogue, with the writer retaining editorial control and agency at every step.

4. Evaluation with Professional Writers and Performance Stagings

Dramatron’s effectiveness was evaluated in a user study comprising 15 professional playwrights, screenwriters, directors, and improvisers in 2-hour compensated co-writing sessions. Each participant:

• Authored a log line and collaboratively built a script stepwise with Dramatron.
• Completed a nine-question post-session Likert survey assessing helpfulness, collaboration, ease of use, creative expression, surprise, output ownership, and other factors.
• Underwent a qualitative interview regarding workflow, limitations, and creative impact.
• Had their input/output quantitatively logged via character-level Levenshtein distance, lemma-based Jaccard similarity (novelty preference), and repetition metrics.

A subset of five co-written scripts was staged at the 2022 Edmonton International Fringe Festival (“Plays By Bots”) by live performers, followed by professional reviews and post-performance discussions. Reviewers commented on the world coherence, emergent robotic humor, and the creative interplay between generated dialogue and actor improvisation.

5. Key Findings and Thematic Analysis

Quantitative and qualitative results highlight both the practical utility and limitations of Dramatron in professional settings.

• Collaborative Utility: 84% of writers found Dramatron helpful, 77% reported a sense of true collaboration, 77% could express their creative goals, and 92% were surprised by its responses.
• Novelty Preference: Lemma-Jaccard scores indicated preference for suggestions lexically dissimilar to the original log line, evidencing value in model-driven novelty.
• Editing Behavior: While dialogue and character lists were often substantially rewritten (high Levenshtein distance), scene outlines were typically lightly curated (low absolute edit distance). Manual pruning of degenerate loops was routine.
• Workflow Strengths: The explicit hierarchical pipeline made overall narrative shape salient, allowed structural issues to be fixed early, and provided a “palette” of outputs at each stage.
• Creative Limitations: Noted shortcomings included lack of true commonsense or embodied reasoning, “on-the-nose” dialogue lacking subtext, persistent gender stereotypes, and inter-scene inconsistency due to parallel generation. The reliance on a well-crafted log line was sometimes discordant with playwrights’ organic, theme-first approaches.
• Perceived Ownership: Only 46% of participants felt pride in the final script, aligning with the tool’s positioning as ideation aid or dramaturgical collaborator rather than sole author.

6. Participatory Design, Ethical Considerations, and System Refinement

Throughout development, Dramatron’s prompts were refined using real-time expert feedback from professional users—a participatory design cycle consisting of incremental prompt tweaks (e.g., inserting previous-beat context, simplifying plot seeds). This process rapidly improved usability and output coherence.

Ethical concerns raised independently by writers include:

• Bias and Stereotypes: Outputs sometimes reflected stereotypes or misogyny present in the pretrained corpus.
• Plagiarism: Concerns about the origins of training data and unintentional copying arose.
• Creative Labor Displacement: Acknowledgment that co-creative tools may shift economic value away from formulaic tasks historically performed by humans.

Mitigations include maintaining a human-in-the-loop for vetting outputs, encouraging plagiarism checking, and framing Dramatron as an “artist support network” or dramaturgical partner, not as a standalone author.

A plausible implication is that the adoption of hierarchical, prompt-chaining pipelines and participatory design approaches can increase both the controllability and creative acceptability of machine-generated long-form texts, while also bringing unique design and ethical trade-offs that must be addressed in real-world use (Mirowski et al., 2022).