An editable writing scaffold improved brief quality and later prompts
In plain words
When AI helps write a brief, people still need to decide what they want the brief to say and how to build it. This study tested a writing interface that turns those decisions into editable goal fields and task-specific planning controls.
In two simulated professional writing tasks, essays scored higher on average when the scaffold was available than with the same chat-and-canvas interface without it. The study also found a higher average score on a later task after the scaffold was removed, but only tested short-term transfer.
The research
When AI produces more of a written task, people still need to decide what they want, organize the work, and judge the result. The study asks whether an interface can support those decisions and whether any writing advantage remains once the interface’s support is removed.
The study tested a dynamic writing scaffold that turns goals and task breakdowns into editable controls. It compared writing with and without the scaffold, then assessed prompts in a later task after its removal.
The authors tested whether an interface could help writers set goals and organize work when using AI to draft professional briefs. Their dynamic scaffold made goals explicit, then offered editable, task-specific planning controls.
In a randomized study, participants wrote one brief with or without the scaffold and a second brief without it.
GPT-5.4’s average essay-quality score was higher for the scaffold group in both tasks. In the second task, prompt sequences also showed clearer goals and more task breakdown after scaffold removal.
These results support short-term gains in writing quality and prompting, while participants reported greater load and less helpfulness during scaffold use.
The study tested two simulated writing tasks in one online session. The later recall result came from 159 follow-up respondents and counted identifying a topic, not understanding or remembering the essay’s content.
Analysis
Deep
- Original paper authors
- Viktor Kewenig, Lev Tankelevitch, Sean Rintel, Advait Sarkar, Abigail Sellen
- Submitted to arXiv
- License
- arXiv non-exclusive