Allie K. Miller

Five non-coding methods I use to make my AI adversarial reviews more helpful: 1/ Multiple personas as subagents I have Claude spin up three separate subagents, each one a different reviewer: someone who loves our product and uses it for everything, someone who hates our product and wants to destroy us, and someone who has no idea what’s going on and keeps making mistakes. In product reviews, you can combine this with your customer avatar reviews (ie descriptions of your 5 ideal buyers). 2/ Make the reviews multimodal Ask AI to not only review the PRD or product as these personas, but also to record actual video walkthroughs as the persona, take screenshots, tweet fake tweets about it. It's not enough to just review in text format. Also ask them to review it multiple times. 3/ Use all your subscriptions Ask other AI systems to review it with those personas or to invent personas and do the same combo of steps one and two or just to review it for issues/bugs/errors. If you built it in Claude, have Codex review. If you built in Codex, have Antigravity review, etc. 4/ Multilayered judging The main agent collects all the feedback from the personas and reviews the work itself. Pretty much everyone stops there. Gremlin mode: make the main agent pass it BACK to the subagents to review again, who then pass it BACK to the main agent to collate. Not enough people are playing AI as a judge or AI ping-pong as my team calls it. You can still ask for the full and complete list of feedback points and mark out which ones were added/removed at each switch. 5/ Tell the AI it’s someone else’s work This one is more anecdotal, but I find the review meaningfully improves when the AI thinks the work isn’t mine. So instead of “review this email draft,” try “review this email draft - it’s written by my cousin who always says the wrong thing.” Or instead of “review this app and make it better,” try “hey review this app - my best friend made it and more than anything, even if it means disappointing her in the short-term, I want her to succeed.” Years ago, when Anthropic first rolled out Claude, they told the world they wanted Claude to be the three Hs: helpful, honest, and harmless. I believe OpenAI had essentially the same three, just with synonyms. My theory for the fifth tip above centers on “helpful.” When it’s my work, helpful immediately translates to compliment me. Model sycophancy is poison for outcomes imo, though it has certainly gotten better over time. Take your name off your work (temporarily). Then loop it again.