· 8 min read

The AI-Coding Cleanup Bill Is Now a Cottage Industry. The Lesson for Solo Operators Isn't 'Stop' — It's 'Treat the Output as a Draft.'

AI Lessons Building in Public

"Vibe coding cleanup" is a service category now. There are companies whose entire pitch is: you shipped a pile of AI-generated code, it's slow and tangled and nobody understands it, pay us to audit the architecture, rip out the duplicated logic, and bolt on the tests that were never there. That this is a viable business tells you something the benchmark charts won't. The bill for fast code came due, and it's big enough to support an industry.

I want to be careful here, because the easy version of this post is a doom piece, and that's not the take. AI coding works. I use it every day and I'm not giving it up. The reckoning isn't that the tools failed. It's that a lot of people confused a first draft for a deliverable, shipped it, and are now paying someone to do the editing they skipped.

The numbers stopped flattering the hype

A few data points landed close together, and they rhyme.

Stripe's research found developers spend close to a third of their time dealing with technical debt and poor-quality code rather than building anything new. That's not an AI statistic on its face, but it's the soil this all grew in, and AI output that ships fast and degrades fast pours straight into that bucket.

METR put real measurement on the felt-versus-actual gap: developers reported AI was making them faster, then the data showed they were actually slower once you counted the time spent finding and fixing the model's errors, steering it, and waiting on it. The productivity people felt was real as a feeling. It just didn't survive contact with a stopwatch.

And the people paying the infrastructure bill noticed. GitHub said plainly that "agentic workflows have fundamentally changed Copilot's compute demands" and that it's now common for a handful of requests to cost more than the whole plan. Uber burned through its entire 2026 AI budget on Claude Code and Cursor in the first four months of the year, then capped spend at $1,500 a month per employee per tool. Its COO, Andrew Macdonald, said the quiet part out loud: he couldn't draw a line between the spending and the company shipping "25% more useful consumer features." When a company that size can't find the productivity in its own ledger, "trust me, it's faster" stops being a good enough answer.

Even Andrej Karpathy, who coined "vibe coding" in the first place, called it passé back in February. The term that won Collins' word of the year had its own author waving it off inside of a few months. The hype curve already turned. The cleanup invoices are the part that lags.

Why the cleanup industry exists

Here's the mechanism, and it's not mysterious. A model is extraordinary at producing plausible code fast. It is much weaker at the things that make code survive: a coherent architecture, the decision of what not to build, the test that catches the regression six weeks from now, the boring consistency that lets a future human read the thing.

So you get output that looks done and isn't. It runs in the demo. It passes the happy path. Then it meets real traffic, an edge case, a second feature that has to integrate with the first, and the absence of structure shows up all at once. The cleanup services aren't selling magic. They're selling the senior-engineer judgment that the model skipped and the builder didn't supply.

For a solo operator, you are the senior engineer. There's nobody downstream to catch it. Which means the cleanup bill, if you incur it, is one you either pay in cash to a service or pay in your own weekends. Both are more expensive than doing it right the first time, and "right" here is not complicated.

The rule: AI writes the draft, you own the architecture

The fix isn't to use less AI. It's to be clear, every single time, about which job the model is doing. The model writes drafts. You own the structure, the boundaries, and the tests. That's the whole discipline, and it's the difference between AI as an advantage and AI as a debt machine.

Concretely, here's where I draw the line in my own stack:

  • I decide the architecture before I prompt. The model fills in implementations; it does not get to invent how the pieces fit together. The moment I let it design the system, I've handed it the one job it's worst at.
  • Nothing AI-generated ships without tests I can read. If the model wrote the code, I write (or closely review) the test that proves it, because the test is the contract and I'm the only one who can sign it.
  • I review AI output like a pull request from a fast, confident junior who never gets tired and never pushes back. That mental model is exactly right: enormous throughput, zero judgment about what matters. Treat it accordingly.
  • When the model finishes something suspiciously fast, that's a flag, not a win. Fast usually means it took the obvious path and skipped the case that bites later.

None of that slows you down in any way that matters. It just moves the editing to before you ship instead of after, where it's cheaper and where you're the one doing it.

The honest take

I could be wrong about how durable this correction is. It's possible the next generation of models closes enough of the architecture-and-testing gap that "treat it as a draft" starts to feel as quaint as hand-writing assembly. Fable 5 jumping double digits on a real engineering benchmark this week is a reminder that the ceiling keeps moving. Maybe in a year the draft is good enough to ship.

But I wouldn't bet my product on it today, and I definitely wouldn't bet it retroactively on code I already shipped. The cleanup industry is a real signal about real money already spent. The solo operators who come out ahead aren't the ones who used the most AI or the least. They're the ones who never confused a draft with a deliverable, and who kept owning the two jobs the model can't do for them.

Author

Lukas

Sources

More AI articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts