I never approve implementation until the plan that AI made survives interrogation.
The pattern:
1️⃣ Build the implementation plan by AI in detail.
2️⃣ Review it. Ask hard questions. What happens when the input is null? What is the retry behaviour? What does this look like in six months?
3️⃣ Probe the assumptions. Work through the edge cases.
4️⃣ Only when it holds up does it get approved.
This is not a prompt. This is a design review.
The AI is the sounding board. The questions you ask reveal what the plan is missing. Gaps surface before any code exists, not at 11pm when something breaks in production.
And when a design choice is genuinely unclear, implement both options. Compare real code. An afternoon with AI, not a week as before.
Real code tells you things a whiteboard does not. Hidden complexity in option A. Testing is easier with option B.
For architects this is next level decision making, not on hunches but on real proof.