Intent

Every session begins the same way: with a sentence that says what should be true when the work is finished. The sentence matters more than the tool. A precise intent is a contract between you and the agent — it defines the shape of success before a single file changes.

Write the intent as an outcome, not an instruction. “Add validation to the checkout form” names a surface; “Reject invalid payments before they reach the order service” names a condition the system must satisfy. The second form gives the agent something to verify against, and it gives you a way to know when the work is actually done.

The best prompt is the one you could hand to a careful stranger and watch them return with the right change.

Keep the intent small enough to hold in memory. When a task has several moving parts, break it into a sequence of intents and let each one complete before the next begins. Agents reason best about one coherent goal at a time.

Finally, say what you do not want. Naming the boundary — the files that must not move, the tests that must stay green, the style that must be preserved — is as important as naming the destination. Intent is a map with the roads you refuse to take drawn in.