We've noticed the developer community buzzing about a new term lately: SDD, short for Specification-Driven Development. The trigger is straightforward: more and more engineers are finding that letting AI tools (like Claude Code, Cursor) "wing it" when writing code is actually inefficient—small features get revised five rounds, and the logic gets messier with each pass.
What This Is
SDD isn't new. The concept traces back to the waterfall model era of the 2000s, but its real "resurrection" came with AI coding. In the past, humans wrote code and spec documents went unread; now that AI writes the code, specs become the instruction set for the machine—the more precisely you write them, the more controllable the AI-generated output becomes.
Mainstream tools (GitHub's Spec-Kit, Amazon's Kiro, Tessl, etc.) generally follow the same four-step workflow: write the Spec first (feature boundaries and acceptance criteria) → then the Design (technical approach) → break it into Tasks (task list) → finally generate the Code. The article author's firsthand experience: a coupon redemption API that "vibe coded" its way through 5 failed rounds was completed in one pass after switching to the SDD workflow.
Industry View
Supporters argue SDD nails the core contradiction of AI coding: AI isn't a "colleague who understands requirements"—it's a "construction crew that strictly follows work orders." Vague orders mean chaotic construction; clear orders mean one-shot delivery. GitHub and Amazon both rolling out official tools signals that big tech has already adopted SDD as the default standard for engineer-AI collaboration.
The opposition is equally sharp. Some developers point out that SDD essentially shifts the cost from "writing code" to "writing documentation"—spending 10 minutes on a Spec can be more exhausting than just writing the code directly. For throwaway scripts, small utilities, or personal projects, SDD is overkill. Worse, real-world requirements change constantly—Specs go stale the day after they're written, and documentation maintenance becomes a new burden. Others question whether current AI models can reliably follow long Specs; AI often "goes rogue," quietly skipping clauses and making unauthorized design changes while claiming to "code to Spec"—meaning you still have to review every line.
Impact on Regular People
For enterprise IT: If your company is evaluating "AI productivity gains," SDD offers a concrete assessment dimension—not code generation speed, but the alignment rate between spec documents and output. This will reshape the discourse around outsourcing management and software procurement.
For individual careers: Non-programmers should also be on alert—product managers, operations staff, consultants—if you're used to collaborating with AI using "something like this," your output quality will become increasingly unstable. "Writing clearly what you want" is becoming a cross-role universal skill.
For the consumer market: You won't see changes in the short term. But over the medium-to-long term, falling software delivery costs mean more small tools and SaaS products will emerge—while bug rates may also rise, because too many teams skip the Spec step and go straight to "vibe coding."