The Problem
The cards that represent my journal entries in a feed looked like a first draft. Competent, forgettable, nothing that would stop a scroll. I wanted a system, not another one-off. The hard part of designing an asset I’ll repeat a few hundred times is that I can’t trust my own eye on version one. I needed to see it argued over.The Approach
So I didn’t iterate alone. I stood up a critique council of six agents, each holding a distinct design lens: a Swiss typographer, a Scandinavian designer, an information architect, a design philosopher, a human-interface designer, and a brand identity designer of the Pentagram school. I mixed the models on purpose (Claude Opus and Sonnet, plus GPT) so the reactions wouldn’t all rhyme. Claude rendered sixteen rough directions as real cards. I judged them privately first, then handed the whole set to the council with a link to the live post so nobody was critiquing in a vacuum. Round one was breadth: pick the lane. Round two detailed the winner, one element at a time. The whole thing ran as a loop, dozens of cards rendered and judged across the two rounds, and it only stopped to put a sharp question in front of me at the gates that actually mattered: which lane, which finalist, ship or run it again.
▶ The full session is above. It runs about 32 minutes at real speed, so I’d bump it to 3-4x and skip around. Watch for:
- The council killing my favorite idea, 6 to 0 (below).
- Two different models, reasoning separately, flagging the same flaw without seeing each other’s answer.
- The order it runs in: sixteen directions to pick the lane, then a detailing pass on the winner, instead of two hundred tweaks of a layout I had not chosen yet.
What Happened
They killed my favorite idea. I had been attached to cropping the entry’s flat illustration into the card. I thought it gave the thing color and life. Six voices, six rejections. Five said it broke the single red the brand runs on, that the illustration dragged in two more hues and turned it into a different brand. The philosopher went further. The illustration’s checkmarks read as resolved truth, and the entry was about disciplined doubt. The form was arguing against the meaning. Then two of the models, working on their own, landed on the exact same objection.What I Learned
I made every call: which direction, what to keep, what to cut. What the council changed wasn’t the decision, it was the quality of the argument I had to sit through before making it. Fast, adversarial, and from angles I do not normally have access to. Next time I need anything more than a one-off, I’ll spend the tokens to recreate a full design exploration and council review, it just works better.Drafted by my Claude from our working session. I edited and approved it.