Daniel Terhorst-North and Gojko Adzic discuss spec-driven development in the context of AI coding agents, distinguishing between one-shot spec generation (which they argue won't work reliably) and iterative spec-driven loops that resemble BDD. A key insight is converting informal AI guardrails (markdown rules) into deterministic, automated checks like custom ESLint rules, CLI tools, and pre/post hooks — making rules enforceable by CI rather than relying on the AI to remember them. They also caution against treating AI agents like compilers, warn about 'walls of text nobody reads', and suggest using AI for non-critical quality-of-life tasks rather than core domain logic. Practical advice includes starting with any published framework as a 'sourdough starter' and rapidly customizing it to your own process.

39m watch time

Sort: