Reusable AI Assistant Contract

Reusable AI Assistant Contract

Contract fields

A reusable assistant should have one named owner, one primary user, one recurring job, and one decision boundary. Its contract records acceptable inputs, authoritative sources, output shape, prohibited uses, permissions, review, failure response, maintenance, and retirement.

Job and success

Write the job as a transformation that can be observed. "Turn my approved notes into a podcast-outline draft that follows the current Venture Step structure" is testable. "Be a world-class podcast expert" is not.

Success should include quality and reviewer effort. A fast draft that requires a full rewrite has not saved useful time.

Evidence and context

List what the assistant may use and what it must not infer. Mark sources by authority, version, and freshness. If current information is required, name the approved retrieval path or require escalation.

Authority

Grant the least capability needed. A drafting assistant may prepare text but not publish it. A reporting assistant may summarize an authorized export but not access a live mailbox or send a report unless separate permission exists.

Failure and escalation

Define what the assistant should do when the input is missing, sources conflict, instructions conflict, private data appears, or the requested action exceeds permission. Useful behavior includes asking a narrow question, returning a partial result with a visible gap, or stopping for a named reviewer.

Maintenance

Record the instruction version, source versions, model or product surface, test set, owner, review date, known failures, and retirement trigger. Every material change should reopen evaluation.

Safe starting scope

The E033 note-organizing and outline-drafting examples are appropriate starting points when the inputs are authorized and the output remains a draft. Travel, finance, health, legal, employment, security, and external communication need stronger evidence and ownership.