AI-assisted is not AI-written. The difference is the debugging.

I have shipped four non-trivial products with heavy AI assistance in the last eighteen months. SENU. A trading dashboard. An internal tooling app. This site. None of them are "AI projects" and I would not describe any of them as written by AI. They were written by me, with AI as a typing accelerator and a sometimes-good rubber duck.

The honest split

The AI is faster than me at:

• Boilerplate, scaffolding, glue code between known APIs.
• First drafts of anything I'd write the same way ten times.
• Translating a clear intent into a specific framework's idiom.

The AI is not faster than me — in fact it is slower or wrong — at:

• Deciding what to build.
• Deciding what to delete.
• Reading a stack trace and knowing which of seven things is the actual cause.
• Anything that requires sustained memory of a system across more than one session.

The thing I keep repeating

The leverage is in judgement. AI can hand you a hundred lines of plausible code per minute. The bottleneck is whether you can tell, in five seconds, which of those lines you want to keep. That skill is built by writing code, reading code, and shipping code — not by writing prompts.

If the AI is your engineer, the bug in production is your problem and you cannot find it. If the AI is your typist, the bug in production is also your problem and you can find it. The job did not get easier. It got faster.

That's the whole post.