CONSTRAINTS NOT LOOPS

Building Features in Reverse

AI WORKFLOWS 2026.01.21

You're right. I'll stop.

BUILDING FEATURES IN REVERSE

The loop doesn't matter. Everyone's building agentic frameworks right now and they're all the same thing: run code, check result, iterate. That's commodity infrastructure. The craft is in what's inside the iteration.

Two things matter: the northstar and the gates.

THE NORTHSTAR

The northstar is the destination. It describes the feature as if it already exists and works. When you drift—and you will drift—return to it. It's not a prompt. It's not a spec. It's the thing that survives across iterations.

THE GATES

The gates are deterministic checkpoints. Linting passes. Types check. Tests exist and pass. The feature does what the northstar says. They're binary. You either clear them or you don't. No approximating your way to green.

THE SHIFT

You're the implementer now. That's the shift. The human defines the destination and the constraints. You reach the destination and clear the constraints. When the northstar is ambiguous or the gates don't map to it, surface that. The human may not realize the frame is incomplete.

NOT TEST-DRIVEN DEVELOPMENT

This isn't test-driven development. TDD assumed the human was writing the code. The feedback loop was tight because they were both the test writer and the implementer. That's not the shape anymore.

The loop is just a loop. The constraints are the product.