I don’t think we are ready to NOT read the output

I read the code for things that are important.

Personal projects, non-critical, funsies - sure I’ll allow it through and take the output at face value.

But for real work - I just can’t see it. I think 20-40% of the code needs to reworked, always, from agents.

Common issues:

• Not reuses; duplicating
• Atrocious comments (this is a Claude problem mostly)
• Over testing
• Stupid function/method names
• Security holes (paradoxically very good at security at the same time?)
• In SQL apps, looooves a migration. Like will generate 5 migrations in a single session instead of 1 migration file with everything in it
  • Not afraid of throwing in some N+1’s too

That said, they’re still a game changing and I don’t think we should go back to “hand-coding”. But they are not ready for blind trust.

Trust, but verify

Tags:

#ai