There is a methodology getting attention on Hacker News today called Verified Spec-Driven Development (VSDD). 130+ points, dozens of comments, and for once the discussion is actually useful.
I read the full spec. Here is what is genuinely good about it — and the one thing builders moving fast should steal immediately.
What VSDD Actually Is
VSDD fuses three things you already know:
- Spec-Driven Development: Write the contract before writing any code. The spec is the source of truth.
- Test-Driven Development: Tests come before implementation. Red → Green → Refactor.
- Adversarial Verification: A separate AI "reviewer" tears apart the spec and code looking for holes.
The key move: these are not competing philosophies but sequential gates. You cannot move to tests until the spec is airtight. You cannot move to code until the tests exist. You cannot ship until an adversarial reviewer fails to find real flaws.
The human is the Architect — strategic vision, domain expertise, final authority. The AI is the Builder. A second, separate AI instance is the Adversary.
What Is Actually Good Here
The spec-first discipline is the real insight — and it is not new, but AI makes it newly enforceable.
When you write code yourself, you can be vague about requirements because your brain fills in the gaps. When you hand a task to an AI agent, vagueness is fatal. The agent fills gaps in the worst possible way: confidently, quickly, and wrong.
Forcing yourself to write a behavioral contract first — preconditions, postconditions, edge cases, failure modes — is not overhead. It is the work. The implementation is almost mechanical after that.
This matches what the best AI-native builders I know actually do. They do not type "build me a payment system." They write three paragraphs about what the payment system must do, what it must never do, and what happens when it fails. Then they hand that to the agent.
The output quality difference is not marginal. It is an order of magnitude.
The Adversarial Reviewer Trick
The part I had not seen formalized before: using a separate AI instance with fresh context as a hyper-critical reviewer.
The reasoning is sound. The Builder AI has context about why decisions were made. The Adversary AI has none — it only sees the output. This is closer to how a real code review works: a colleague does not care about your implementation journey, only whether the result makes sense.
In practice: write your spec and implementation with Claude. Then paste it into a fresh Gemini session with the prompt "You are a hyper-critical senior engineer. Find every flaw, ambiguity, and missing edge case. Do not be polite."
The results are genuinely useful. It catches things you missed precisely because you did not miss them — you made a deliberate choice the Adversary correctly identifies as unjustified.
What to Skip If You Are Moving Fast
The full VSDD pipeline — Epics → Issues → Sub-issues ("beads"), formal verification strategies, mathematical proofs of properties — is built for teams shipping production systems that cannot fail.
If you are a solo founder validating an idea, that overhead will kill you.
Steal the spec discipline. Skip the ceremony.
Concretely:
- Before any AI coding session, write 5-10 sentences: what this feature does, what it never does, what the 3 most likely failure modes are.
- Hand that to your AI agent as context.
- When it produces something, paste it into a fresh AI session and ask for the harshest possible critique.
- Fix what is real. Ignore what is theoretical.
That is 80% of VSDD value in 20% of the effort.
The Real Unlock
What VSDD points at — even if the full methodology is too heavy for most solo builders — is the shift from "can the AI write code" to "can I tell the AI exactly what to build."
The bottleneck is no longer implementation. It is specification.
The builders who win in the next two years will not be the ones who prompt best. They will be the ones who think clearest about what they are actually trying to build — and can write it down precisely enough that an agent has no room to improvise badly.
That skill is worth more than any framework.
Building AI products for SMBs in LATAM. Spec-first approach: rooxai.com
Top comments (0)