Amateurs use AI as a search box. Operators use it as staff. The difference is the brief: a role, a scope, an input, a deliverable, and a definition of done — the same things you'd give a contractor, because ambiguity produces the same garbage in both cases.
A useful delegation has edges: 'You are reviewing this text ONLY for factual claims. Don't touch style. Output a numbered list of claims with a confidence note each.' Scope keeps the model from wandering; the deliverable spec makes output checkable.
The seed is a fully-scoped delegation. Run it on two models and grade them like employees: who stayed in scope, who followed the deliverable spec, who freelanced beyond the brief.