A feature idea is not yet an implementation target. Specification authoring turns it into one through structured dialogue between people and a coding agent. The resulting specification records intent precisely enough to guide implementation and verify the outcome, rather than leaving the agent to infer intent from a single prompt. section 2.2 section 2.2.1
Spec-driven development separates that dialogue into three focused phases before implementation:
- Discover establishes the business context: what the feature is meant to do, its purpose, and its scope. section 2.2.1
- Design addresses technical choices, including the technology stack and architecture, using the discovered intent as its basis. section 2.2.1
- Task turns the preceding artifacts into a task list that guides the agent's execution. section 2.2.1
This order is not a requirement to specify an entire application upfront. Features are specified incrementally, and their specifications can be revisited when implementation reveals something new. During execution, the specification remains the agent's target; afterward, it becomes durable context that future features can build on. section 2.2.1 section 2.2
When one engineer becomes the scribe
A product specialist describes a feature, an architect supplies constraints, and an engineer writes both contributions into the specification. If the agent then asks about business behavior, the engineer must answer from second-hand understanding or return to the specialist. Direct participation by the people who hold each part of the intent avoids making that translation bottleneck the specification's source of truth. section 4.1.2
Specifications can also point readers and agents to relevant code, tests, security guidance, and coding standards. Kept accessible, they reduce the need to reconstruct context for each new feature; kept behind a single scribe, they risk preserving an incomplete translation instead. section 4.0 section 2.2 section 4.1.2

