Debugging LLMs: turn vague failures into reproducible fixes. 1) Reduce to a minimal input + pin model+samp params (temp, top_p, seed). 2) Encode the bug as an automated test asserting concrete invariants (regex/fields/hash). 3) Attach prompt+response diff to the issue. 🔧 #LLMDev