Pair Programming With AI: How to Talk So It Actually Understands

The bottom line: AI is an obedient pair partner with zero common sense

The biggest illusion in AI pairing is treating it like a human. With a human you can say "refactor this bit" — they know what "this bit" means, what 'good' looks like, what not to touch. AI has none of that. It executes literally, and loves to 'helpfully' do extra. So the first principle of AI communication: say out loud everything you assumed was common sense.

Habit 1: state the goal before the approach

Bad: "Rewrite this function with async/await." Good: "Users report the button freezes for 2 seconds on page load (goal: eliminate the freeze). I suspect a sync request in this function blocks the main thread — verify that first, then rewrite with async/await if confirmed." The difference: the first turns your diagnosis into a conclusion, and if you're wrong, the AI will flawlessly execute a wrong plan. The second gives it a chance to verify — and having AI verify your hypothesis first is the cheapest way to avoid rework.

Habit 2: one task at a time

"Also write the tests, update the docs, and fix that bug while you're at it" is a recipe for disaster. AI attention is a finite resource; quality degrades exponentially as tasks pile up. The right posture is a task queue: one thing, verify, then the next. It sounds slower but is much faster, because rework drops off a cliff. A useful sentence: "Do only X for now, stop and wait for my confirmation — don't start Y."

Habit 3: acceptance criteria instead of vibes

"Make it elegant" or "make it faster" means nothing to AI. Translate into verifiable criteria: "function under 50 lines," "p99 latency under 200ms on this endpoint," "tests cover all three branches." Whatever you can write as a test assertion is what AI can understand as a requirement. Acceptance criteria have a bonus: afterwards, have the AI check its own work against them — a free extra review round.

Habit 4: know when to shut up

Many people pair with AI like chatterboxes: interrupting every three lines. The result is fragmented output from constantly broken flow. Better rhythm: give full context on big tasks, then let go; check in at milestones, not mid-step. Concretely: front-load background, constraints, and acceptance criteria at task start; intervene mid-task only when the agent asks or visibly drifts; review everything at the end. AI doesn't need your real-time company — it needs a clear starting line and finish line.

The one-line summary

Pairing with AI is about making explicit everything in your head that felt "obvious": goals, constraints, acceptance criteria, task boundaries. Get these four right and the same model performs like a different machine in your hands — the difference was never the model; it's how you talk to it.