Why AI builds collapse at seventy percent

The first screen is always beautiful. The trouble starts at prompt eleven, when the auth you built does not fit the data model you never decided.

Every AI build starts the same way: a gorgeous first screen and a rush of momentum. Then, somewhere around prompt eleven, it quietly falls apart.

The pattern we kept seeing

The first screen is always the easy part. What separates the products that ship from the ones that stall at seventy percent is whether the hard decisions were made before the prompts started — auth before billing, the data model before the features.

We wrote this down because we kept giving the same advice. Decide the shape, then let the builder fill it in. The order is the product.

What we changed

Once the plan is explicit, every later prompt inherits it. Nothing drifts, nothing contradicts prompt three by prompt eleven, and when a decision changes the affected prompts are flagged — never silently rewritten.