The idea
A reliable assistant starts with a narrow task, explicit inputs and a defined output. State what the assistant must not do and when it must ask for clarification. Instructions guide a model but are not a security boundary: sensitive actions need application-level permissions and human approval.
Worked example
A fictional assistant drafts replies to shop delivery questions. It receives an approved policy excerpt and an anonymized question. It returns a draft and the policy section used. It must not invent delivery dates, issue refunds or send messages. Missing order details trigger a clarification request.
Try it
Write a specification for one drafting task. Include the input, output format, two forbidden actions and a clarification rule. Use a fictional customer question to check whether your specification is complete.