What Breaks When LLMs Turn Fictional Prose into Screenplays

See how to convert fictional scenes into screenplays, preserving subtext and intent. Learn a workflow to transform creative input into structured output.

ChatGPT Plus GPT-based LLM workflow Prompt Engineering Iterative context scaffolding Human editorial review

Overview

I built a human-in-the-loop workflow that converts a narrative fiction scene into screenplay/teleplay format while trying to preserve subtext, pacing, character power dynamics, and scene function. For the demo, I’ll show a short prose scene, the weaker output from a basic prompt, and then the improved screenplay-style output after I changed the prompt structure and context constraints to better control exposition, visual action, dialogue compression, and tone.

Tech stack

ChatGPT Plus

A professional-grade subscription providing priority access to GPT-4o, advanced data analysis, and custom GPT creation.

ChatGPT Plus delivers peak performance for power users through guaranteed access to OpenAI’s most capable models (GPT-4 and GPT-4o) even during high-traffic windows. Subscribers get 5x more message capacity, early access to experimental tools like Sora or Voice Mode, and the ability to build tailored GPTs for specific workflows. Whether you are executing complex Python scripts via Advanced Data Analysis or generating high-resolution DALL-E 3 imagery, this tier ensures faster response times and a comprehensive feature set for technical and creative tasks.

https://chatgpt.com/premium

View projects
GPT-based LLM workflow

A structured pipeline that chains prompt engineering, retrieval-augmented generation (RAG), and fine-tuning to automate complex reasoning tasks.

Modern GPT workflows move beyond simple chat boxes to integrate vector databases like Pinecone and orchestration frameworks like LangChain. These systems use recursive loops and multi-step reasoning (Chain of Thought) to handle high-stakes data processing. By connecting a GPT-4o model to real-time APIs and proprietary datasets, operators reduce hallucination rates below 2% and automate 80% of manual documentation tasks. The focus is on precision: using system instructions to enforce output schemas (JSON mode) and implementing automated evaluation scripts to maintain performance at scale.

https://platform.openai.com/docs/guides/optimizing-llm-accuracy

View projects
Prompt Engineering

Prompt Engineering is the discipline of structuring inputs (prompts) to Large Language Models (LLMs) to reliably and efficiently elicit a desired, high-quality output.

This is the core skill for maximizing performance from models like GPT-4 and Claude 3: it's the art and science of guiding an AI. The process involves systematic iteration and applying specific techniques to control the model's behavior and reduce 'hallucination.' Key advanced methods include Chain-of-Thought (CoT) prompting, which forces the LLM to process complex problems step-by-step, and Few-Shot prompting (providing 2-3 examples) to establish a clear output format or style. Mastery of these methods directly translates to tangible gains: improved accuracy, reduced API costs from fewer retries, and production-ready outputs for applications like customer service bots or code generation.

https://www.promptingguide.ai/

View projects
Iterative context scaffolding

A prompting framework that builds complex LLM reasoning through layered, incremental information delivery.

Iterative context scaffolding replaces massive, single-shot prompts with a structured sequence of logical steps. By feeding the model specific constraints and domain data in stages (often 3 to 5 distinct phases), operators reduce hallucination rates and improve output accuracy for technical tasks. This method mirrors human pedagogical scaffolding: start with core definitions, layer in edge cases, and finalize with execution parameters. It is particularly effective for large-scale code refactoring and complex legal analysis where context windows often become saturated by irrelevant noise.

https://arxiv.org/abs/2303.17580

View projects
Human editorial review

Expert human oversight ensuring accuracy, nuance, and ethical alignment in content production.

Human editorial review serves as the critical final filter (the 'human in the loop') for high-stakes publishing and AI-generated outputs. While LLMs handle scale, human editors apply 15+ years of domain expertise to catch hallucinations, verify 10-K filings, and ensure brand voice consistency. This process involves rigorous fact-checking, legal compliance checks, and sensitivity reads that automated systems consistently miss. By integrating experienced journalists and subject matter experts, organizations maintain a 99.9% accuracy rate and protect their reputation against the risks of unvetted algorithmic content.

https://www.nytimes.com/2023/10/13/insider/the-human-touch-in-the-age-of-ai.html

View projects

What Breaks When LLMs Turn Fictional Prose into Screenplays

Related talks