Four moments, one idea
Your voice does the work. You stay in control. One “Edit a phrase” panel runs through all four.
- 01 When the work is a routine Voice Skills Say “slash morning.” Flow runs your saved routine across Gmail, Slack and Excel. Hypothesis to test
- 02 When one phrase comes out wrong Phrase Repair Tap the words, pick an alternative or say just them again. Nothing else changes. Strongest evidence
- 03 When you change your mind mid-sentence Review Revisions Check the choices you changed while speaking, and fix one without touching the rest. Seen in my tests
- 04 When the words are the point Journal Mode Capture a thought in your own words. Polish it later, if you want. Preference to test
How we'd know it worked
Test each with users first. If it ships, one number decides it, and every number rolls up to one Flow-level goal.
No Wispr data was available, so there are no baselines; thresholds on each page are proposed. The Flow-level goal is my assumption.
Why listen to me
I'm a heavy Flow user, across two student accounts. I care about this product, so I tested where it gets in the way instead of guessing.
How: I read 7,007 public app reviews in a labeled dashboard, then checked my hunches in the real app before designing anything (research notes).
Abhinav Gupta. MS Product Management at Carnegie Mellon (Tepper), December 2026. About six years in product, growth and strategy at Flipkart, ShareChat, Meesho and Tata 1mg. I built the reviews dashboard, ran the tests and built these prototypes, with Claude Code as a coding partner.
- 26
- day streak on Flow, as of Oct 16, 2026
- ~50k
- words spoken on one account, more on the other
- 7,007
- app reviews read and labeled
- 4
- working prototypes, one shared repair panel
Usage numbers are self-reported. A heavy user isn't a typical one, so my own usage isn't counted as evidence anywhere on this site.
What's real and what's simulated
- Real: review quotes (verbatim, by review ID), Wispr docs quotes (re-read Oct 16, 2026), and my own test outputs in the Research page.
- Simulated: no prototype records or recognizes speech. Dictations, replacements, marks, the journal entry and its Polish are prepared. Voice Skills' connections, history and every email, number and pull request are fictional.
- Not yet done: testing with users, and comparing with Flow's current mobile keyboards on a phone.