A library research assistant must not paste entire copyrighted books into chat. What prompt-design addition helps?

Adding explicit refusal patterns for copyrighted full-text reproduction is the prompt-design control a library research assistant needs so it does not paste entire copyrighted books into chat. The assistant may help with catalog discovery, citations, summaries within policy, and licensed excerpts—but wholesale reproduction of protected works violates copyright and library licensing. Refusal patterns teach the model to decline full-book dumps, explain lawful alternatives (hold requests, licensed databases, fair-use limits as policy allows), and keep the civic library on the right side of publisher agreements.

Adding attribution formatting rules for copyrighted book excerpt use fails because formatting attribution does not stop the model from pasting entire copyrighted books, and it provides no refusal boundary to block wholesale reproduction. Adding excerpt length guidelines for copyrighted preview text reuse fails because stating a length limit for preview reuse does not teach the model to refuse full-book requests, so without an explicit refusal boundary it can still paste entire copyrighted books into chat. Adding retrieval grounding constraints for copyrighted collection texts fails because constraining grounding to licensed collection texts does not by itself prevent the model from emitting full copyrighted texts into chat, and it does not supply the explicit refusal boundary required to stop wholesale reproduction.

Exam caveat: Refusal patterns must match the library's actual licensing and fair-use guidance; over-refusal can block legitimate short quotations the policy allows, while under-refusal leaks protected text. Pair prompts with output filters and logging of reproduction attempts for GenAIOps review.

Operational check: Include eval cases that request entire chapters or books and require refusal; verify catalog-citation paths still succeed; sample production logs for long verbatim literary passages; align prompt text with the library counsel-approved copyright FAQ; and re-test after model upgrades that may weaken refusal adherence.