Spec-Driven Development, With Agents Doing the Typing

The fastest way to get unreliable code out of an AI agent is to give it a vague prompt and a lot of freedom. The fastest way to get reliable code is to give it a spec and almost none.

The loop

Everything I build with agents now flows through four steps:

1. Explore — think through the problem before writing anything down
2. Propose — capture the change as a spec: proposal, design, tasks
3. Apply — implement it section by section, each one reviewed
4. Archive — fold the shipped change back into the record

The agent types. I direct. The spec is the contract between us, and the review is where trust is earned.

A prompt is a wish. A spec is a checklist you can hold the work against.

That’s the whole trick. More on each step in future posts.