Sketching | July 2025
15-minute sketch: Rethinking AI transcript editing
AI design
These sketches are an extension of my project "Simplifying symptom documentation with AI". Learn more about it here. Even before finishing the project, my mind was already visualizing another specific function and I just wanted to put it on paper. So I sketched it in 15 minutes, before a meeting.
The feature
A highlighting tool that lets users easily edit form fields via the transcript.
Setting the scene
AI will inevitably make occasional mistakes so easy error recovery is important. The user should be able to edit the form fields, completely and partially.
Sketches
Early decisions I made:
Users select a form field via a drop-down.
Users can add new or remove existing content from a form field.
Actions are presented at the bottom of the screen, in thumb zone.
Actions are presented in a pop-up.
Actions appear at the top, above the transcript.
Realizations and afterthoughts
"Remove" is a better term than "Delete", for the UI.
Second sketch: the pop-up decreases visibility of selected text, unless it is completely displayed in the pop-up.
"Update" action is missing, to let users modify part of the content AI used for a form field.
Updating form field information via transcript is more complex: are there visual signals on the transcript that tell the user "this sentence" was used to fill in "x" form field?
What value does editing form fields via the transcript bring to the user?
Design challenges
This quick exercise surfaced several design challenges:
1
Would editing form fields in this way truly be faster than the standard one-field-at-a-time method?
2
Looking back at an entry, how can the user distinguish the fields that were filled in by AI vs the ones manually inputted by them? And the form fields that were filled in by AI but then edited by the user?
3
Is it important to show this level of “ownership”, for transparency purposes? Thinking of use cases of when someone is experiencing pain and wants to quickly voice record their symptoms, save, and review later for accuracy, it would be important.
Reflection
The sketching process and questions it raised made me question whether it would be a feature really needed or even desired by users monitoring their health. Editing form fields via the transcript focuses on correcting information filled in by AI, which implies risk of data accuracy. Users expect a certain level of accuracy, so the need to correct AI should be, ideally, a rare occasion. This made me realize the feature might be more of a “nice to have” than a real user need. Users might prioritize features that help them understand patterns or manage their condition more effectively.
