Verified Spec-Driven Development (VSDD)

Concerns about VSDD/TDD with AI

• Writing tests first implicitly invents an API; with an AI “test writer,” this can lead to hallucinated, unstable interfaces that later tests merely distort rather than improve.
• Several commenters report AI-produced code + tests that technically satisfy specs with high coverage but form an unmaintainable ball of mud; extensibility and resilience under change are underemphasized.
• Token waste is a recurring issue: models tend to rewrite entire files for small edits or loop on partial changes, driving up cost.

Specs vs Exploration and Iteration

• One camp argues you can’t meaningfully spec systems you don’t yet know how to build; with code cost near zero, you should favor rapid exploration: many agents, many variants, keep only the good parts.
• Others counter that a spec is about “what,” not “how”: you can and should specify desired behavior even before knowing implementation details, and that formal or semi-formal specs are powerful design tools, not just verification.
• Many stress that specs need not be fully up-front “waterfall”; they can be iterated alongside implementation, serving as a stable reference for checking AI output.

AI-Assisted Workflows and Tools

• Described workflows include:
  • Human-steered SPEC.md + PLAN.md, iterative steps gated by human review (“LLM as junior dev”).
  • Using AI to draft high-level design, then humans refine, then AI implements tests and code.
  • Static call-graph tools to give models a concise structural view of the codebase.
  • External orchestration/guardrail systems (e.g., TDD frameworks and workflow engines) that force models through deterministic steps rather than trusting in-prompt discipline.

Testing, Verification, and Fundamental Limits

• Debate over TDD vs BDD: many note that common testing styles already look like BDD; others warn that tests AIs can easily generate are also tests they can game.
• Some highlight that verifying properties of programs is inherently hard (model checking, P vs NP); any process claiming to “solve programming” must hide trade-offs in where certainty is relaxed.
• Formal verification is held up as the only unfoolable verifier, but acknowledged as costly and only practical where specs are far simpler than implementations.

Skepticism and Social/Process Observations

• Multiple commenters believe the VSDD writeup is AI-generated “slop” or “word salad” and refuse to engage without concrete case studies (real specs, real bugs caught).
• There’s discomfort with AI-written prose in technical discourse: perceived as disrespectful and low-effort compared to the reader’s investment.
• Others see spec-heavy, language-centric workflows as appealing mostly to people who prefer talking over coding, and warn that much of this can become elaborate procrastination.
• Some report that AI has shifted the bottleneck from coding to requirements discovery, sponsor engagement, and timely feedback—regardless of methodology.