Spec-driven development has become the dominant pattern for working with AI coding agents in 2026. Tools like GitHub Spec Kit, now past 111,000 stars and adopted across more than 30 coding agents, and AWS Kiro’s EARS-notation requirements both treat the spec as a living, structured source of truth an agent references continuously — not a document written once and set aside while the “real work” happens in the code.
Spec-driven product management takes that same discipline and applies it one step earlier: to the decision about what to build, before there’s a spec to write at all.
The parallel, made explicit
Spec-driven development exists because prompting an AI coding agent from scratch every session produces inconsistent, context-blind output. The fix was to give the agent a persistent, structured object to reference — a spec that survives across sessions. Product decisions have exactly the same failure mode: a PM re-explaining “why are we building this” in every sprint planning because the reasoning was never captured anywhere durable. The fix is structurally identical — a persistent, structured decision that survives across meetings, quarters, and reorgs.
What a spec-driven product decision actually contains
Borrowing the same rigor spec-driven engineering uses, a spec-driven product decision should include: the evidence that justifies it, with source and freshness attached; the specific claim being made and the trade-offs considered; an explicit non-goal, so scope doesn’t silently expand later; and a way to check, after the fact, whether the outcome matched the intent. Most PRDs today have the first two. Almost none have the last two.
Why this needs to be product-owned, not just engineering-owned
It would be easy to assume spec-driven engineering tools solve this by themselves — if the engineering spec is well-structured, isn’t that enough? It isn’t, because engineering specs start from an already-made decision. Spec-driven product management is what produces that decision in the first place, grounded in evidence a PM gathered, not a technical constraint an engineer discovered. Skipping this step means the engineering spec, however well-structured, is formalizing a guess.
How this shows up in Argus
This is the operating model behind Argus end to end: gather evidence, sharpen it into what’s supported versus what’s still in conflict, propose a decision with its trade-offs explicit, get it reviewed by the right approver, and carry it into delivery in a form engineering can query directly. It’s spec-driven development’s discipline, applied to the question spec-driven development assumes has already been answered.