Coding with an agent tends to settle into the same loop: prompt, check, re-prompt and repeat, burning time and tokens each pass. There's a better way, and it starts before any code gets written.
This talk puts two approaches head to head: the prompt-and-iterate loop most of us have fallen into, and spec-driven development. You'll see what each one produces, and what a spec gives you that prompting can't, including decisions you can review before any code exists and context your whole team can build on.
