80% of My Vibe Coding Was QA, Not Vibing

by•

I built a production app entirely through an agent. The "vibe" part—describing what I wanted—was maybe 20% of the work.

The other 80% was the same discipline as any build. Catching a timezone off-by-one. A file upload limit rejecting real photos. Missing auto-capitalization. Refusing to accept "it's fixed" without proof.

Vibe coding didn't skip the engineering. It relocated it. Instead of writing code, I was writing precise instructions and doing QA like a hawk.

Has your split matched that, or did you find a different bottleneck?

9 views

Add a comment

Replies

Best

On the upload limit: our support form takes up to 3 screenshots of 2 MB each, so the clients shrink every image to a JPEG of at most 1600 px before sending, and the server checks the file type from its first bytes and ignores the type the client declares. On "it's fixed" without proof: our agent rules make proof part of the task. After a change that could fail to compile it runs the analyzer, after a feature lands it installs the build on the Android emulator and drives it, and since the debug build lags there, it takes a screenshot before each tap rather than tapping blind. A build that fails is treated as information, not as a mistake to hide.

Same split here, and refusing "it's fixed" without proof is the rule I ended up writing down. My global CLAUDE.md is 562 lines now, and one rule is: don't ask me to check anything before you've looked at the real screen yourself. A hook even stops the agent from claiming a fix it hasn't verified. I have 94 hooks in total, so the QA didn't go away, it moved into hooks and E2E tests.