Prompt AI to be "verifiable": principles of prompt design

The core of using AI well is to ask in a verifiable form. Andrej Karpathy said in 2026 that "traditional computers automate what you can specify in code, and LLMs automate what you can verify." So a prompt gets more accurate the more it makes the output verifiable. ASAP organizes this principle into prompt design.

The one line that upends prompt conventional wisdom

Most prompt advice has stayed at the level of expression: "be specific," "assign a role," and so on. Karpathy's claim shifts the focus from expression to gradeability. The problem stops being how to phrase a prompt well and becomes how to design it so you can judge whether the result is right or wrong. From this angle, even a finely tuned prompt is untrustworthy if there is no bar to tell a correct answer from a wrong one.

Why verifiability decides capability

Where an LLM is strong versus weak is split by verifiability. Models are trained by reinforcement learning in environments where answers can be graded (tests, scores, proofs), so they spike in capability on such tasks. Asking in a verifiable form draws out that strength and lets people catch the result too. In effect, the way a model is trained dictates the way it should be used. The closer your problem resembles the environment the model learned in, the more fully its capability shows.

Three design principles

A verifiable prompt is built from 3 design principles, in the spirit of Karpathy's point. Apply them in order.

State the correctness bar — pin down what counts as "right" with tests, examples, or a format.
Require output in a verifiable form — ask for checkable output like code-plus-tests, a table, or a checklist.
Require a self-check step — have it review and look for counterexamples before giving the answer.

All three are devices that make the result checkable.

Bad prompt vs good prompt

The same task is split by verifiability, with LLM output diverging sharply. The following contrast is the key.

Bad prompt	Good prompt
"Fix this function"	"Make this input produce this output, and write that case as a test too"
"Write marketing copy"	"Output each of the 5 items below as one sentence, in a table"
"Analyze this"	"Give the conclusion in one line first, then the basis as 3 numbers"

The right column draws out the model's strength and makes the result verifiable. What all three share is that they plant a bar for judgment into a vague request. With the left column you cannot tell from the result whether it worked; with the right, pass or fail is immediate.

Handling tasks that cannot be verified

Tasks that are hard to verify are risky under any prompt. With no way to grade the answer, the model can produce a plausible wrong answer that people miss too. Such tasks should be split into small verifiable pieces, or have human judgment explicitly inserted. For practitioners, this makes the principle a matter of decomposition skill. Rather than handing over inherently ungradeable work like strategy calls or taste, the people who use AI safely are the ones who carve it into verifiable subtasks.

The limits of this principle

Verifiability is not a universal yardstick. The cost of setting up a correctness bar can exceed the work itself, and qualitative judgment that resists being expressed as a test still belongs to people. Even so, the claim that a good prompt is a verifiable prompt holds. Stating the correctness bar, requiring output in a checkable form, and asking for a self-check are how you practice it.

Source: ASAP synthesis grounded in Andrej Karpathy, "From Vibe Coding to Agentic Engineering" (Sequoia AI Ascent 2026; specify-able to verify-able automation).