Andrej Karpathy shares the ultimate prompt philosophy: talk to AI about anything for 10 minutes with voice chat

OpenAI co-founder and current Anthropic pre-training team member Andrej Karpathy posted on X on Tuesday (7/21) a fixed routine for collaborating with LLMs. When he encounters complex tasks, he doesn’t type; he switches directly to voice input and talks continuously for 10 minutes. The content can jump around, be chaotic, and have no rhyme or reason. His observation is that the model is very good at reconstructing what you truly want to do from long, rambling incoherence. The version it spits back is usually cleaner than what you would have said yourself.
(Background: OpenAI co-founder Andrej Karpathy announced that he is joining Anthropic—returning to the very front lines of LLM research)
(Background add-on: Andrej Karpathy distilled the “CLAUDE.md four major principles,” sparking a GitHub wave and pushing the accuracy rate for AI-written Code past 90%)

Too lazy to type—just talk. On Tuesday (7/21), Andrej Karpathy wrote on X about one fixed routine for collaborating with an LLM. When there’s too much background to explain and his fingers can’t keep up with his brain, he leans back against the chair, switches to /voice voice input, and then speaks continuously for 10 minutes. “It’s completely a mess—anything can be said, full-stream-of-consciousness the whole time.”

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10…

— Andrej Karpathy (@karpathy) July 21, 2026

Sometimes he starts by greeting the model and adds a line at the beginning like, “Switched to speech recognition—sorry in advance for any typos.” Other times, he turns this monologue into a kind of mini interview, letting the model ask a few rounds of follow-up questions, digging out the scattered clues one by one.

Voice input itself isn’t new. What Karpathy is really talking about is the amount of input. Instead of spending effort organizing your ideas into a neat set of instructions, you simply throw the whole bundle of thoughts over and let the model take it apart, recombine it, and reassemble it.

Say more up front, change less later

His reasoning is very straightforward. “For some reason, LLMs are extremely good at reconstructing long, incoherent monologues. The echo they give you is often cleaner than the jumble of thoughts you originally had.” he wrote. The result is that mind meld (mind fusion—meaning a gradual alignment of understanding of each other’s intentions between a person and the model) improved, and after that point there were actually fewer places he needed to go back and correct.

This line highlights the real cost structure of this approach. Most people save the effort of typing at the beginning; the trade-off comes later, with repeated revisions and back-and-forth messages to clarify what you mean. Karpathy’s method is to put all the cost upfront into those first 10 minutes: using one super-heavy dose of background information to replace what would otherwise be five rounds of “That’s not what I mean.”

Those who oppose it worry about the same thing

The replies area on X is broadly split into two camps. Supporters say this method already accounts for 80% of their prompt, and some people even describe that after switching to voice as the main input, output efficiency increases by 5 to 8 times. Multilingual users also like that it doesn’t care about the language.

The voices against it focus on the same concern: outsourcing the whole act of “making yourself clear” to the model—long term, will your own organizational ability and writing ability degrade? Some people propose a compromise: once the model has organized everything, you then manually distill it again yourself—don’t take it all at face value.

In reality, neither side is entirely right or wrong. Karpathy treats the model as a mirror that can organize—not a ghostwriter. The only difference is whether you read back the version it produces after it’s organized.

**