Karpathy says voice rambling gives LLMs better context than a short prompt

JeremyNguyenPhD · x · 2026-07-22

Andrej Karpathy’s workflow for working with LLMs is to speak through a problem out loud first—sometimes on a walk or in the car—then feed the transcription to the model.

The core idea is to ramble freely in voice mode for about 10 minutes.
Let the transcription capture tangents, corrections, and changing thoughts.
Karpathy says the extra context helps LLMs reconstruct the real task better than a short typed prompt.
The post frames this as a practical way to give models more useful bits without having to organize everything perfectly upfront.

Original post →

Karpathy says voice rambling gives LLMs better context than a short prompt

More from Apps