Bootstrapping XML schema definitions with Claude Opus 4.6: A case study (the good, the bad, and the ugly)

Great post - very relatable.

LLMs aren’t failing loudly, they’re failing convincingly. The constant generator, bypassing self-hosting, even deleting tests - all classic “looks correct” over “is correct.”

Your TL;DR is spot on: without a strong test harness, vibe coding breaks.

Feels like the real skill now is designing systems where the AI can’t cheat.