Why does talking through an idea produce better results than typing it?
Think out loud: using voice as the front end to AI
Typing forces you to have the thought before you write it. Speaking doesn't — and for the messy early stage of an idea, that's the point.
There's a stage of an idea where you know something is there but you can't yet say what. Typing is a bad interface for that stage. The keyboard imposes a finished sentence, and imposing a finished sentence on an unfinished thought usually kills it or narrows it prematurely.
Speaking doesn't have that property. You can be circular, contradict yourself, trail off, and restart — and all of it is still usable input.
Why the messiness helps
A rambling five-minute explanation contains more of your actual thinking than a tidy three-line prompt. The digressions are information. The place where you say "well, sort of, except..." is often exactly where the interesting part is, and it's precisely the bit you'd have edited out while typing.
Handed that transcript, a model has enough to do something genuinely useful: reflect back the structure you didn't know you had.
The loop
- Talk for a few minutes without editing. Explain the idea as if to a colleague who's interested but has no context.
- Ask for structure, not polish. "What am I actually arguing? Where did I contradict myself? What did I assume without saying?"
- React to what comes back. The value is in noticing "no, that's not quite it" — which is far easier than producing the right version from a blank page.
- Only then ask for prose. And only if you need prose.
The part that surprised me
The most valuable output is usually not the polished writing. It's the question the model asks back — the one that reveals you'd been avoiding a decision. That only happens if the input was messy enough to contain the avoidance, which is the thing typing quietly edits away.