Should factual content and behavior rules live separately?

Fin's Blueprint Team proposes a rule for keeping agent instructions from growing unmanageable: "if it's about content or information, it belongs in the knowledge base. If it's about tone or how to handle specific situations, it belongs in your Agent's instructions." The team frames the alternative as a known failure mode — "Teams keep adding rules every time a new edge case comes up. Before long, the LLM has paragraphs of instructions to process before it can respond" — and reports having both caused and observed this elsewhere: "I've seen this happen at Fin and I've heard the same from other teams. The instinct is always to add more, but the discipline is knowing when to stop."

The reasoning given is economy of instruction, not capability: a short directive like "Be direct about pricing" does "more than a paragraph explaining the philosophy behind your pricing communication strategy," on the premise that factual lookup and behavioral judgment are different kinds of information that belong in different places — one retrievable on demand, one always loaded into the prompt. The team's suggested audit is to find conversations where the agent "answered correctly but the customer still escalated or left negative feedback," on the logic that correctness failures point to the knowledge base while dissatisfaction despite correctness points to the instructions, then to trim instructions that have "grown beyond a few focused rules" by moving content out of them into the knowledge base.

The rule gives a practitioner-side name to a split that Why do language models engage with conversational distractors? documents from the model's side: that paper finds models reliably follow what-to-do instructions but not what-to-ignore instructions, meaning they handle content directives better than behavioral constraints. Fin's content/behavior split functions as a workaround for that asymmetry — by keeping behavioral rules few and explicit in the instructions and offloading content volume into a separately retrieved knowledge base, it avoids asking the instruction-following mechanism to hold the kind of large, accumulating rule set that paper's distractor tests show degrades compliance.

The excerpt gives no mechanism for why mixing content and behavior rules degrades an agent beyond the team's own experience of "paragraphs of instructions to process," and no measurement of how many rules or what prompt length triggers the degradation it warns against. It is a heuristic for information architecture in system-prompt design, not a tested claim about instruction-following limits — but it converges with instruction-tuning research on exactly where behavioral instructions are hardest for models to hold, which strengthens the case that the split tracks a real asymmetry rather than a stylistic preference.