Here’s a question Andrej Karpathy uses to expose what today’s models still can’t do: I want to go to a car wash 50 metres away. Should I drive or walk? State-of-the-art models say walk — it’s close. They miss the obvious: you need the car at the car wash. I tried it on four frontier models after hearing this; they all said walk.
The point isn’t that models are dumb. It’s that they’re brilliant at what can be measured and blind to context they were never given. Which is why, in Karpathy’s framing — distilled here from austin.marchese’s walkthrough of his AISN 2026 talk — almost everyone is prompting wrong.
The fix is the spec, and Karpathy is pointedly unimpressed with the lightweight version most tools ship. On plan mode: “I don’t even like the plan mode… I think there’s something more general here where you have to work with your agent to design a spec that is very detailed.” (Translated nothing here — he said it in English; the channel is one of the few that does.)
Two moves make a spec real:
Uncover the goal, not the task. “Create an end-of-month report” is a task. The goal is the decision that report drives — the conclusion you’re trying to reach. A model can execute the task but can never decide the goal; that information lives in your head. His trick for getting it out: tell the agent to interview you first. The spec is how your understanding gets into a form the model can use.
Work agile, not waterfall. The natural failure mode with capable agents is to hand them the whole thing at once and wait for a finished product. Karpathy’s counter is to break the work into small buckets and check direction throughout. People over-trust agents precisely because the agents are good — and waterfall hides the drift until it’s expensive.
The honest caveat: “write a better spec” is advice that’s easy to nod at and hard to operationalize. The interview-me technique is the one concrete, stealable habit here — and it costs you nothing to try on your next non-trivial task before you write a single line of prompt.
Source: austin.marchese — Stop Prompting Claude. Use Karpathy’s Method Instead (YouTube), summarizing Karpathy’s AISN 2026 remarks.