When I hand AI a real piece of work, the task itself is rarely the hard part. The hard part is getting the AI to judge its own output the way a genuine expert would. Left alone, it produces something that looks fine and lands generic.
So before it starts, I have it build a tribe of mentors for that kind of work, and go.
The idea is simple. I tell the AI what I am trying to do, whether it is writing code, planning a campaign, or fixing an operations problem, and I ask it to assemble the people and frameworks that are genuinely good at that specific work, then to judge its own draft the way they would.
The point is not the names. It is the questions those mentors would ask. For a landing page, a useful bench might check three things: does the page make the one important action obvious, does it answer the visitor’s biggest objection before asking for anything, and would a first-time reader know what to do in five seconds? Good marketing gets judged differently than good code, and that is the whole value. The AI stops reaching for a generic idea of “good” and starts measuring against the standard that fits the work.
Try this on your next real task. Paste this into your AI: “Before you start, build a short mentor bench for this kind of work: three to five experts or frameworks known for it, and the specific questions they would ask to judge the result. Do the work, then grade your own draft against those questions before showing me.” Keep the benches that produce sharper work and reuse them.
Always expect friction.