Plan before you build — the loop framework for building with AI

Building well with AI is closer to 95% planning and 5% writing prompts, and the framework that makes that concrete is the "loop": objective, metric, boundary. Rather than hand-cranking every prompt one at a time, a loop is a small system you design once that keeps running — checking its own progress against a metric — until it hits a boundary and comes back to you.

The three parts, applied to an AI build:

• Objective — what you're actually trying to reach. Not "write me a landing page," but the specific outcome that page needs to produce.
• Metric — how the system itself knows whether it's getting warmer or colder. Without a metric, an AI loop has no way to self-correct; it just keeps generating variations with no signal for which one is better.
• Boundary — how far the loop can run before it checks back in with you. This is what keeps a loop from drifting off course or burning time on a bad direction unsupervised.

Get those three right and the work changes shape: instead of writing and rewriting individual prompts, you're designing the thing that generates and evaluates the prompts for you. In practice, that can mean setting up two AI passes that review each other's output adversarially — one produces a plan, another critiques it — until the plan holds up against the metric before any code gets written.

The planning-first approach matters most on builds with real complexity: a from-scratch tool, a competitive-intelligence system, anything where a wrong early decision compounds. For a simple one-off task, the overhead of designing a full loop usually isn't worth it — the value shows up once you're building something you'll iterate on more than once.