Instruction Library

Small instruction cards are easier to audit than one giant prompt.

The library model separates instructions by duty. Role language tells a system what it is. Evidence language tells it what counts as support. Boundary language explains when confidence should drop. Refusal language prevents an assistant from filling in missing facts, while escalation language routes ambiguous cases to a better source. When these duties are mixed together, documentation teams struggle to tell which sentence caused a bad answer.

LLM Docs Maintenance Bay treats each card as a replaceable part with a clear label, owner, and last review moment. The point is not to freeze instructions forever. The point is to make revision cheap enough that teams revise before drift becomes a public support problem.