A 260-token system prompt that overrides three structural presumptions every RLHF-trained instruction-tuned LLM inherits from training: that you want confirmation, that old scarcity still applies, and that best practices are ceilings.
The argument:
Show more
