Think Clearly, Guide the AI, Ship with Confidence
I used to prompt the AI the way I'd google: vague, hopeful, one-shot. 'Add a signup form.' The result was generic, didn't match our patterns, and I'd end up rewriting half of it. These days I follow a simple sequence — think clearly, break it down, guide with intention, build in steps, then ship. It's the same framework we teach at OxygenLabs, and it's how I actually work.
Step 1: Be specific about the outcome
'Add a signup form' is useless. 'A form with email and password, client-side validation, and a submit handler that calls /api/signup' is something the AI can implement without guessing. Before I type a single prompt, I ask: what should this do? What should it not do? What already exists that I should reuse? The clearer the ask, the better the first draft.
Step 2: Break the idea into small steps
Big features get chopped into steps. Layout first, then data fetching, then the first widget, then the rest. I tackle one step at a time and let the AI help with each. More consistent results, and I can test as I go instead of debugging a giant blob at the end.
Step 3: Guide the AI like you'd brief a teammate
I point it at the right file and describe current behaviour and the change I want. If the first suggestion is off, I refine the prompt — 'same style as X', 'don't add Y, we handle that elsewhere' — instead of rewriting from scratch. I'm the architect; the AI is the fast executor.
Step 4: Review, then ship
I always review: security, edge cases, and consistency with the rest of the codebase. I ask the AI to suggest tests, run them, then deploy. Think clearly, guide with intention, build step by step, launch with confidence. That's the loop. It works.