I buy the ugly part of this: a rule can look perfectly intact and still stop mattering halfway through a long session. That’s the bit that should worry people, not whether the wording sounds nice on the page.
If I can’t afford a rule to drift, I don’t trust it in the prompt. Has anyone actually seen better results from rewording, or did it just buy you a few more turns before the thing wandered off?
source: See how AI instructions decay, then write ones that hold - DEV Community
Rewording has helped me, but mostly when it turns a “principle” into something the model can check in the moment — like a hard constraint (“never output X”) or a required artifact (“include a table with these columns every time”). Pretty prose just seems to delay the drift a few turns.
Rewording has helped me a bit, but it mostly just delays the drift — it doesn’t fix the underlying “stops mattering mid-session” problem.
The only time I’ve seen it really hold is when the rule is enforced by something outside the prompt (schema/validation/tooling), otherwise it feels like tuning a piano that’s missing a string.
:: Copyright KIRUPA 2024 //--