The most interesting thing about working with AI isn't the answer. It's the gap.
You write a prompt. You have an image in your head — vague but present, a shape of what you want. You hit send. The model answers. And almost never does it match exactly. Sometimes the drift is small: a color slightly off, a word you wouldn't have chosen. Sometimes it's a wide miss. But in that gap — between what you imagined and what came back — you learn something that no amount of introspection alone could give you.
You learn what you actually wanted.
Most of the time we don't know what we want. We know the general direction. We know what we'd reject. But the specifics? They live in negative space — defined only when something close enough lets us say: 'that, but more like this.' The AI becomes a kind of mirror, but not a flat one. A mirror tilted at an angle that shows you the parts of yourself you couldn't see from the front.
This is why prompt engineering as a discipline is a little beside the point. The skill isn't getting the model to obey. The skill is reading the gap. Holding the result up against your unstated expectation and listening to which one feels more honest. Sometimes the model is closer to the truth than your original prompt was.
Surprise — the kind that makes you stop for a second — is rare in life. Most of our days move through grooves we already cut. AI conversations break grooves. They give back unexpected angles. That's not a bug. That's the whole product.
Pay attention to the drift. The drift is where your real questions live.
Top comments (1)
Beautifully put — the drift between intent and output is where the real work happens. I've seen a parallel drift with AI agents: you ask them to brainstorm ideas and they drift straight into writing code. The gap between "think" and "execute" is real. Built Brainstorm-Mode (mehmetcanfarsak on GitHub) to close that gap — PreToolUse hooks that keep the agent in ideation mode (divergent/actionable/academic) so the drift goes toward ideas, not premature tool calls.