Why spec-driven AI development beats vibe coding

The first year of agentic coding produced a generation of developers who type a prompt, accept what comes back, and call it shipped. It works fine until the work gets serious. Once the code touches money, customers, or compliance, vibe-coding produces the same thing it always has: a tower of plausible-looking work that nobody can defend.

The failure mode has data now

If vibe coding worked, the numbers would show it. They show the opposite. In a 2025 randomized controlled trial by METR, experienced open-source developers using early-2025 AI tools took 19% longer to complete real tasks from their own repositories while estimating, even after finishing, that the AI had sped them up by about 20%. And GitClear's analysis of 153 million changed lines found code churn, where lines are reverted or reworked within two weeks of being written, on track to double against the pre-AI baseline, with copy-paste rising and refactoring falling.

Neither study is an argument against AI coding. Both are measurements of what happens when generation gets faster while intent and verification stay informal. The METR developers weren't slowed down by the model's typing; they were slowed down by reviewing, correcting, and re-prompting output that didn't match what they actually needed. That is precisely the cost a spec pays down in advance. And churned code is, almost by definition, code that shipped before anyone wrote down what done meant.

What spec-driven actually means

A spec doesn't have to be a 30-page PRD. At minimum it's three things written down:

• The user-visible behavior you're trying to produce
• The constraints the change must respect, like security, performance, data, and compatibility
• A definition of done that lets a second person verify the work without re-deriving the intent

Write those three things and the work holds up to review, whether it came from an AI agent, a junior engineer, or you in six months.

Here's the whole loop in one picture: the spec sits in the middle, and everything else reads from it or writes back to it:

Why it matters more with AI in the loop

People fill in ambiguity from context. Agents fill it in from training data. That difference is load-bearing. If the spec doesn't say "all timestamps are UTC and we never trust the client clock," the agent will guess. Sometimes its guess matches yours. Sometimes it doesn't, and you find the bug in production three sprints later.

Spec-driven development is how you stop relying on lucky guesses.

It also changes what review is for. When the intent is written down, a reviewer, whether human or automated, checks the work against the spec instead of re-deriving the requirements from the diff. That's the difference between review as verification and review as archaeology, and at agent volume, archaeology doesn't scale. The teams drowning in AI-generated pull requests aren't drowning in code; they're drowning in unstated intent.

The Sigao take

We built Cadence on this. A spec is what makes speed safe, so a team serious about agentic development invests in writing specs the way it once invested in CI. The payoff lands exactly where the studies say the cost lives: less churn, less re-prompting, and review that verifies instead of excavates.

Sources

• Becker, Rush, Barnes, and Rein (METR), "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", 2025.
• GitClear, "Coding on Copilot: 2023 Data Suggests Downward Pressure on Code Quality", 2024.