When the crit rules flagged my dark mode, I did not let the agent fix it however it liked.
I asked for three directions, picked the one I could defend in a room, and made the agent prove its work. Two prompts do most of that.
The first stops it from handing me its average:
Give me three genuinely different directions for this screen. For each: the idea
in one sentence, what it trades away, and who it is for. Do not build any of them yet.The second makes it look at what it built before it tells me it is done:
Take screenshots of what you built at 390 and 1280 pixels wide. Compare them with
the reference. List every difference before you fix any.Claude Code’s own best practices put it plainly: give the agent a check it can run, like a screenshot to compare.
And keep the maker away from the judge. Anthropic’s engineers found that separating the agent doing the work from the agent judging it is what improves the result. Studio taught me that years before any of this existed.
Want it as slides? Download the carousel.
The takeaway
Let the machine ask the questions. Keep the answer.