Vibe Coding Has a Pilot Problem

Mode 1: You Pilot, AI Co-Pilots

I had a computer vision pipeline that had grown into an untestable monolith. 2,000 lines, no tests, tightly coupled modules. It needed to be decomposed — but decomposed in a specific order, because some modules depended on others.

I designed an 8-phase migration plan. I decided which modules to split, what interfaces they needed, and what order to tackle them in. I wrote the test strategy. AI executed each phase — fast — but the architecture decisions were mine.

84 commits. 54 tests. 15 days. The code went from untestable to fully covered. AI typed it. I flew the plane.

This is the mode people think of first — the human as architect, AI as builder. It works best for refactoring, system design, and anything where the sequence of decisions matters as much as the decisions themselves.

Mode 2: AI Pilots Phase 1, You Take Over

I needed a pickleball scoring app. State machine, UI, persistence. I gave AI a clear brief: build it.

AI scaffolded the entire thing in an afternoon. Scoring engine, game flow, full interface. One commit. One session. A working product.

Then I got involved. Because the state machine didn't handle how players actually keep score in recreational play. The UX assumptions were wrong in subtle ways that only someone who plays the sport would catch. I refined edge cases, adjusted the flow, made judgment calls about what mattered.

The fastest path wasn't piloting from minute one. It was letting AI take the first pass, then taking the controls once the shape emerged. You don't need to steer the plane during taxi. You need to steer it through turbulence.

Mode 3: The Co-Pilot Knows a Route You Don't

I was building survivorship analysis for a loan portfolio. I knew what I wanted — a way to visualize when loans default over time, not just whether they default. I asked AI for survivorship analysis techniques.

AI came back with Kaplan-Meier curves — a technique from oncology research. Survival analysis. The same math that tracks how long cancer patients survive after treatment, applied to how long loans survive before default.

I wouldn't have found that on my own. Survivorship analysis is common in machine learning, but Kaplan-Meier's specific roots in biostatistics — its elegant handling of censored data, its non-parametric flexibility — that was the co-pilot bringing specialized knowledge the pilot didn't have.

The skill here wasn't prompting. It was recognition. I knew enough about the problem to evaluate the suggestion and say: yes, that's exactly right. A pilot who can't read the instruments the co-pilot is showing them isn't a pilot.

The Real Skill

The pilot/co-pilot relationship in vibe coding isn't fixed. It's dynamic. The skill isn't “always be the pilot.” It's knowing when to fly, when to hand off the controls, and when to trust a suggestion you wouldn't have thought of yourself.

What stays constant is domain expertise. The rancher who knows calving schedules. The lender who knows what makes a loan go bad. The analyst who knows what survivorship analysis should look like, even if they don't know the specific technique. That's the flight plan. Without it, the co-pilot just circles the runway.

That's what makes this a practice, not just prompting. And it's why the bottleneck was never typing speed. It was knowing what to type.