Give the assistant a defensible source.

List authoritative documents, owners, versions and exclusions before asking a chatbot to learn from a website.

Choose one default version and make exceptions explicit; a correct answer for the wrong release is still a failed answer.

Publicly reachable is not the same as approved for import or public answering; confirm the source boundary before uploading.

Explain product operations without granting the assistant authority to perform account changes.

If an answer is missing or contradictory, repair the documentation first; generated wording does not resolve the source problem.

Compare imported content with the approved source list before evaluating generated answers.

Write questions with expected evidence and failure conditions, not just a list of prompts that sound realistic.

Define the reader job, source boundary, unacceptable failures and owner before the first evaluation.

Budget visitor questions and review questions separately; our local calculator totals attempts, not vendor credits or money.