I posted Labelr on Reddit. Someone corrected the title to say "LLM built" it, not me.

Fair enough. Claude Code wrote most of the code. But here's what it didn't do: figure out that Gmail's API has a 15-second watch expiry that silently drops events. Or that launchd on macOS needs a specific plist format or it fails with no error. Or that Ollama's context window is too small for full email bodies so you need to truncate intelligently. Or that users want retroactive labeling but Gmail's history API paginates differently than the watch API.

AI wrote the code. I debugged the platform. Those are not the same thing.

The current state of AI is a multiplier. It amplifies what you already bring to the table. If you can architect systems, debug edge cases, and make real tradeoffs, AI makes you 10x faster. If you can't, it doesn't matter. Zero times anything is still zero.