Spec-driven development puts a precise spec — the intended behavior and its tests — at the center and lets an agent generate the implementation to satisfy it. The human’s leverage shifts from typing code to specifying intent and verifying output, which is why comprehension and review matter more than raw coding speed.
Worked example: write a precise specification of desired behavior first (as prose, tests, or examples), then have the model implement against it and check the result — making the spec, not the code, the source of truth the AI works from. Gotcha: the spec is only as good as its checks: if acceptance criteria are executable (tests, types, evals) the model can iterate to green on its own; if they are vague prose, you are back to hoping the output is right — so the leverage comes from making ‘correct’ machine-verifiable.