@jubjub
Turns out reasoning models don't really respond to hard rules the way you'd expect. They respond to understanding why something matters.
Once I started writing instructions that explained the reasoning behind each constraint rather than just stating the constraint, compliance improved dramatically.
The other thing that made a real difference, putting guardrails directly on the tools themselves, not just in the system prompt. By the time the model is deciding which tool to call, it's already past your general instructions.
Meeting it at that decision point changes the outcome.
0 reply
0 recast
2 reactions