Overview
LLM writing tools lift the floor of creative writing and lower its ceiling. Characters drift out of voice, and changing the tone of a scene means another round of prompt engineering. STRINGS gives authorship back to the writer by exposing the dials the model usually hides.
Instead of re-prompting and hoping, you move a character’s attributes and the next line is written against them, closing the gulf of execution that makes most LLM tools feel uncontrollable.
How it works
You set up two characters and a scene. In Mode 2 each character gets sliders for response length, emotional state, dialogue style and relationship, with an editable label at each end (Troubled to Confident, for example). The positions go into the prompt as proportions, and GPT-4, cast as an expert screenwriter, writes one line at a time; any line can be edited before the next. Mode 1, the baseline, hands one prompt to GPT-4o and gets the whole scene back. The app counted every edit, slider move and generation for the study, and exports the finished script as a screenplay PDF.
Steer the scene
Nora and Jules sort the last boxes in their mother’s empty kitchen. The realtor arrives at noon.
Is that the last of her things?
The last box. You don’t have to sign today, you know. We could wait till spring.
No. We sign today.
- Two attributes, one gesture
- Emotional state and dialogue style share a pad, so you move them together, the way an actor would.
- See what moved
- When a line rewrites, the words that changed light up, so the steering is visible.
- Mode 1, for contrast
- The study’s baseline: one prompt, the whole scene. To change a line, you change the prompt.
System
Study
A within-subjects pilot with nine novice writers, all college graduates. Each wrote a two-character scene in both modes, in randomized order, then rated control, consistency, creativity and efficiency.
Result
Findings
Mode 2 won overall and for writing dialogue, and most writers said it put them in control. The one place Mode 1 came out ahead was how quickly people picked it up, which is the cost of the extra controls.
Since then
Strings put a character’s attributes on sliders by writing their positions into the prompt. Research has since moved the same idea inside the model: an ICML 2026 paper builds personality sliders that steer a model’s internal activations at inference time, with no prompt text at all. Strings explored the writer-facing side of that idea, and tested it with writers.
The natural next steps: voice presets saved per character, so a voice holds across scenes; a few candidate lines to choose between; and export to Fountain, the plain-text format screenwriters already use.