Unit 16: Specification-Driven AI Development

Fast generation changes the bottleneck.

When code can appear in seconds, the scarce resource becomes human review capacity. If the project grows faster than you can understand it, speed has created review debt rather than progress.

Write the target before the implementation

This Unit begins with a specification: purpose, requirements, constraints, out-of-scope work, interfaces, and acceptance criteria. Only then do you break the work into implementation slices small enough to review.

requirement
   -> acceptance criterion
   -> bounded slice
   -> implementation
   -> test / runtime evidence
   -> commit or reject

You will use change budgets to predict the expected blast radius of a task. A request for one formatter should not casually redesign storage, add a framework, and touch twelve files.

The same process works with or without AI. A human developer also benefits from explicit requirements and small verified slices.

Completion evidence

Leave the Unit with a short specification and a verification matrix connecting each important requirement to a test, runtime check, or other observable evidence. Include one case where you stopped, rejected, or reduced a proposed change because it exceeded the specification or your review capacity.

The goal is not to generate the most code. It is to keep the system explainable while it changes quickly.