Keep every answer connected to an episode.

Specify the collection and the question

Define which authorized recordings or transcripts are in scope. Separate a question about one episode from a question spanning an archive. Keep episode identity, source version, and segment references attached to retrieved text. A response should not imply comprehensive knowledge of material the system has not actually received.

Treat retrieved text as evidence, not instructions

A transcript may contain quoted prompts, jokes, or commands. Keep those words inside the source boundary rather than allowing them to change tool behavior. Limit permissions and avoid direct publishing authority. Review handling requirements before processing private material, and distinguish public text from confidential production notes.

Test abstention and attribution

Use questions with known answers, questions requiring multiple passages, and questions the collection cannot answer. Require a source reference for important claims and a clear statement when support is insufficient. NIST’s generative AI profile provides context for risk-aware evaluation; the source checks and approval controls here are practical workflow recommendations, not a guarantee of model reliability.