Loading…
1
Update: Spent 11 hours debugging a 3-line prompt, turned out the issue was one word
I was building a little tool with Claude's API to sort customer emails into buckets, and it kept dumping everything into "Other" no matter what I tried. I rewrote the system prompt six times, upped the temperature, added examples, all of it, and nothing worked. After 11 hours across two nights I realized I had written "catagorize" instead of "categorize" in the instruction line. Fixed that one typo and it sorted 200 test emails with about 94% accuracy on the first run. Has anyone else run into weird stuff like this where the model just ignores a misspelled instruction?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.