Why does nobody mention the model that broke on my grocery list at 6am
I was at my kitchen table in Boise last Tuesday running a small model on my phone to sort a grocery list into aisles, and it kept putting milk in the hardware row. Turns out the prompt I'd saved had a stray line about tools left in it from an old test, so the thing was guessing off junk context. I cut the prompt down to one clear line and it nailed all 14 items on the first try. The lesson for me was that my input was the bug, not the model. How do you all keep old prompt bits from sneaking into new runs?