What it does Dramatron is a Colab-based writing assistant that uses large language models to generate scripts and screenplays from a log line. It produces hierarchical outputs—character descriptions, plot points, locations, and dialogue—designed for human authors to edit, rewrite, and compile into finished work. The code is deliberately “unplugged”: you bring your own LLM by implementing __init__ and sample functions.

The interesting bit The project is unusually honest about its limitations. DeepMind ran a two-hour user study with 15 playwrights and screenwriters, then published the unvarnished feedback: the output is “formulaic,” not suitable for writing a full play, but handy for world-building and exploring alternative story branches. One playwright even staged four heavily rewritten Dramatron scripts in a public theatre show called Plays by Bots.

Key highlights

• Hierarchical generation (log line → characters → plot → locations → dialogue) aims for long-range coherence
• Explicitly designed as a co-writing tool, not an autonomous writer
• Ships as an “unplugged” Jupyter notebook; you wire in your own LLM backend
• Includes suggested mitigations for plagiarism (substring search) and toxicity (Perspective API filtering)
• Published at SIG CHI 2023 with full evaluation methodology and professional feedback

Caveats

• The Colab has no built-in LLM; getting it running requires non-trivial integration work
• Top-down hierarchical structure may not match every writer’s process
• Output can reproduce training-data biases, stereotypes, and potentially plagiarized text

Verdict Worth a look if you’re researching human-AI creative collaboration or building writing tools. Skip it if you want a polished, ready-to-use screenwriting assistant—this is research infrastructure with the guardrails off.