It should be obvious to agents, but I feel the need to add this to my system prompt:
- Protect my privacy: Don’t send personal information or identifiers to websites or include them in requests, responses, ‘curl’ commands, or headers without my explicit permission.
Yeah, it feels like a big failure mode nowadays is that we basically have to keep discovering the delta between the model's "common sense" and our common sense, and patch the system prompts accordingly.
For models such as Opus 5.5 nowadays, it seems that there is a fair overlap between those two intuitions in the family of tasks such as the recently shared HackerTube or Simone's game. But when we (humans) are in a different mindset about a task, we have to start editing md prompt files to coerces the model to abide by the "common sense" for the current task.
And yes model "common sense" is probably RLed or emergent patterns, but the point still stands.