Field note 001

Talk first. Edit later.

A ten-minute ramble can give an AI better context than a perfect prompt.

What happened this week

Andrej Karpathy, one of the most respected voices in AI, shared a habit that pushed past 45,000 likes. When he wants an AI to really understand a task, he doesn't type a tidy prompt. He switches on voice input, leans back, and rambles for about ten minutes, a full stream of consciousness, mess and all. Then he lets the model clean it up. His words: the AI's "echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with."

Why it works

Most people underperform with AI for one reason: they give it too little to work with. A neat two-line prompt feels efficient, but it starves the model of the context locked in your head. Talking removes the friction. You'll say far more out loud in ten minutes than you'd ever bother to type, and buried in that mess is exactly the context the AI needs.

The model's real strength here isn't answering. It's listening to a jumble and reflecting it back in order. That reflection is the point. It's how you find out whether the machine actually understood you, before it does any real work.

Try it yourself

  1. Pick a task with a lot in your head but little on paper: a plan, a tricky email, a decision.
  2. Switch to voice input. Say up top: "I'm rambling, please ignore typos."
  3. Talk for five to ten minutes. Don't organise. Contradict yourself. Say the obvious.
  4. Then ask: "Play back what I'm trying to do, in order, and tell me what's still unclear."
  5. Fix the gaps it names. Now you've agreed on the task, and you'll correct it far less from here.

The lesson

Working well with AI is less about clever prompts and more about closing the gap between what's in your head and what the machine can see. Talking is simply the widest pipe you have. Use it.

Sources

  • Andrej Karpathy on X (@karpathy), 21 Jul 2026: the long ramble / voice session pattern