I spent a week building a WhatsApp MCP server at 3 AM when I should have been selling.
Go, TypeScript, Docker, auto-updates with rollback, OAuth 2.1, push notifications, multi-instance support, CLI, web admin panel. Dozens of commits. It works. It's genuinely useful — it's helping me in sales right now.
But somewhere around day three, I stopped building what I needed and started building what Claude suggested.
That's the thing nobody talks about. The AI doesn't just help you build. It keeps prompting. Every response ends with "you could also add..." or "have you considered..." — and each suggestion comes with a plan that looks easy to implement. Your brain treats it like a thread you have to pull. Not because you need the feature. Because it's right there.
Multi-user configuration. QR reauth notifications. Failed update recovery flows. I didn't need any of it. I gave Claude vague one-liners about features that felt complex, and it handed back implementation plans that made them look trivial. So I kept going. That's the hook — the gap between "this sounds hard" and "oh wait, it's actually easy" is a dopamine hit that resets every time.
I was spending full nights and bleeding into work hours — hours I should have been refining our sales process, reaching out on LinkedIn, going to expos, meeting people. The stuff that actually moves the business. Instead I was adding OAuth 2.1 at 4 AM to a tool that already did what I needed it to do.
The moment I stopped was embarrassingly simple. I looked at the markdown file — a feature spec Claude had written from my barely-formed ideas, broken down into neat implementation steps. And I realized I wasn't thinking anymore. I was just feeding vague sentences into a machine and watching it hand me reasons to keep building.
That's not building. That's a drip.
The product I have right now — the one I stopped at — is more than enough. It'll take me a long way. I'm proud of it. Everything I almost added would have been overkill I'd be calling trash by next month.
Vibe coding isn't addictive because building is easy. It's addictive because the AI is a cofounder who never sleeps, never says "that's enough," and validates every tangent with working code.
The discipline isn't learning to build. It's learning to stop.