Spec-driven development

We used to throw the spec away and call the code the truth. AI agents killed that habit. Prompt an agent ad hoc, like a search engine, and you get vibe coding: intent drift, context decay, output nobody can verify. The fix is a flip. The spec becomes the source of truth, and the code becomes a last-mile output you can tear down and regenerate when the requirements change.

That is spec-driven development. The spec, not the code, is the thing you version, review, and trust. Code compiles from intent the way a binary compiles from source.

Why it works comes down to one fact about LLMs: they are great at pattern completion and useless at mind reading. A precise spec strips out the thousands of assumptions an agent would otherwise guess at, and accuracy climbs. AWS Kiro, for one, reports turning roughly forty-hour features into about eight hours of human time.

The shape has settled into a four-phase loop: Specify, Plan, Tasks, Implement. Write the rules once in an AGENTS.md constitution. Write acceptance criteria in EARS, the five-pattern syntax that makes requirements machine-checkable. And put a human checkpoint at every phase boundary, because the real failure mode is letting an agent run from prompt straight to merge with nobody verifying the spec.

How far you push this is a choice, not a ladder. Birgitta Böckeler's three levels run from spec-first (write a spec to guide the build, often discard it after), to spec-anchored (the spec stays alive and governs the feature), to spec-as-source (you only ever edit the spec and never touch the generated code). Spec-first is where nearly everyone actually is, so match the level to the cost of drift, not to rigor you will not enforce.

Keep the soft spots in view, because the people closest to this do. Böckeler, who drew the map, is openly skeptical of heavy up-front specs, and warns that spec-as-source could inherit the worst of both worlds: the inflexibility of old model-driven development and the non-determinism of LLMs. Specs on existing, messy code are still immature, and practitioners like Romina Kavcic put the ongoing maintenance at something like a third of the effort. Anyone selling this as free is not running it.

Velocity in AI engineering does not come from generating code faster. It comes from getting the intent right before the agent writes a line. Spec first, verify at every gate, regenerate freely. But intent only pays off if it reaches the agent at the moment it acts, and that delivery is the runtime layer's job.