"Mostly Works" Is Not a Business
I had an idea. Then I made the exact mistake I now warn my clients about.
I used an AI assistant to pressure-test the concept — which is a genuinely good use of these tools. It poked holes, it stress-tested assumptions, it helped me decide the idea was worth pursuing. Then, at the end of the session, it cheerfully offered to just build the whole thing.
I said yes. Of course I did. It's fast, it's confident, and it makes you feel like the hard part is behind you.
The hard part was not behind me.
The first crack
The first sign that something was off — and honestly, with 20+ years of coding behind me, I should have caught it sooner — was a bug in the posting feature.
On-demand posting had a problem. I flagged it, and the AI fixed it instantly. Great.
The same problem existed in scheduled posting. I flagged that too. And this time it couldn't fix it. It couldn't even find it. Same symptom, same behavior, and it was completely stumped.
That doesn't happen if there's one well-structured piece of code doing the work. So I asked the question experience trains you to ask:
"Are you using the same code for both?"
It wasn't. There were two separate code paths doing the same job. I fixed one; the other quietly rotted. I solved that one — and then the next issue showed up. And the next. Each fix surfaced another seam that should never have been there.
Then I asked the questions that actually matter
Bugs I can fix. Bugs are the easy part. The questions that decide whether you have a product or just a demo are the ones about money:
- What does this cost me per user?
- What are the hosting fees?
- Storage?
- API calls — and what happens to all of the above at scale?
Silence.
The AI had handed me something that mostly worked and had no idea what it would cost to run. Which meant I had no idea what to charge for it — or whether I'd lose money on every single customer I signed up. For a business, that's not a rough edge. That's the whole ballgame, unanswered.
Why this happens (and why it isn't your fault)
Here's the part nobody in the AI-build world likes to say out loud, because people assume it means they're bad at using AI:
You didn't fail at AI. You outgrew the platform.
Vibe-coding platforms are extraordinarily good at one thing: getting you to "mostly works." A demo. A proof. A first version you can put in front of a real person. That is real value, and it's the point of those tools.
But "mostly works" is exactly where their vision ends. They're blind to duplicated logic, because a demo doesn't punish it. They're blind to what breaks at scale, because a demo never gets there. And they're blind to unit economics, because a demo has no bill at the end of the month. None of those things show up in the thing they're optimized to produce. All of them show up in production.
Hitting that ceiling doesn't mean you did it wrong. It means the thing worked well enough to matter now — and "matters now" is a different engineering problem than "works in a demo."
The rescue
So I did what I've done for my entire career. I took the code the AI gleefully supplied and ran it through the process I've honed over two decades — the same process I use on any codebase that has to survive contact with real users.
I put the project on a tried, true, and upgradable stack. I collapsed the duplicated paths. I made the costs visible and measurable, so I know exactly what a user costs me and exactly what I need to charge. I built in the room to grow that the original never had.
Now it's a bulletproof program with known, measurable costs, providing a service I can confidently put in front of clients and sell.
That project is Show Your Work. It exists as a real, sellable product today because it got rescued — not because it got generated.
If your build is stuck at "mostly works"
If any of this sounds familiar — the demo that dazzled, the fixes that break two other things, the sinking realization that you can't answer "what does this cost" — you're not at a dead end. You're at the exact point where AI stops and engineering starts.
That's the work I do: taking vibe-coded and half-built projects and turning them into production software that ships and holds up. Twenty-plus years of it, no junior team, no agency markup.
If your build is stuck at "mostly works," that's not the end of the project. That's where I come in.