Spec-driven development

Spec-driven development flips the chat-first way of using coding agents. You do not steer the model through a chat and hope. You write the plan down first. What should change. Where the edges are. How you will know it worked. The agent then builds against that document. The code is judged against the spec, never against a memory of the chat.

The practice grew as an answer to drift. Drift is code that looks right but quietly solves the wrong problem, because nobody wrote the goal down. Tools followed. GitHub's Spec Kit, AWS Kiro, and most agent products now support some spec, plan, build loop.

A spec does not need to be long. Here is the useful test. Could a second engineer, or a second agent, pick it up cold? Would they make roughly the same change? If yes, the spec is doing its job.