stage · Talking to it
19Talk, don't type
Out loud you give it three times the context you'd type
There's a trick almost nobody uses, even though it's free and works from the first minute: don't type the request. Say it out loud.
Here's why it works. When you type, you save effort. You shorten, you drop the details, you write three lines instead of ten. And details are exactly what the model needs - without them it fills the gaps by guessing.
Out loud, one minute gets you what five minutes of typing would. And you say the parts you'd normally be too lazy to type: why it has to work this way, what you already tried, what worries you, what it should look like in the end.
Andrej Karpathy - one of the OpenAI founders - described his own version: he leans back in the chair, switches on voice input and rambles for ten minutes without policing his words. Sometimes he opens with a line like: switching to speech recognition, sorry for the typos in advance.
And here's the part that surprises people: models are very good at reconstructing meaning from a rambling mess. What comes back is cleaner than what you said - your own thoughts, combed straight.
So don't be shy about umming and repeating yourself. A voice message to a friend you'd re-record three times, because all those 'well, like, anyway' bits are embarrassing. The model doesn't count your filler words. It doesn't care.
Where to find it: the ChatGPT and Claude apps both have a microphone button, and on a phone the keyboard's own dictation works fine. Start with a task you'd normally put off because it's tedious to explain.