By Friday afternoon, I’d reached the seventh version of a quote

I was doing personal-brand consulting for a client in Hangzhou. Every time, they said, “Make it feel more premium,” so I revised, then rejected my own work. At 4 p.m., I stared at a screen full of feedback and realized the problem wasn’t a lack of effort. I wasn’t turning each round of feedback into evidence for the next attempt.

I got stuck that day too. At one point, I even considered asking AI to reply to the client directly. What I actually needed wasn’t more generated answers, but a clear view of feedback, action, and result.

The useful part isn’t the terminology

Reinforcement learning often sounds distant. At its core, it’s a loop: make a choice, see the result, receive a reward or penalty, then adjust slightly next time. Say I post a short video and one client comments, “Useful.” I explain more about that topic next time. If nobody responds, I try a different opening.

I later drew three columns in a Feishu doc: “What I did—What they said—What to keep or change next time.” My friend Lin Xiaonan, a freelance photographer in Guangzhou, uses a similar table to review her session-booking conversations. When clients prefer natural light, she schedules the next park shoot for the morning. When a client asks for extra retouching at the last minute, she clarifies the number of included delivery rounds beforehand.

This doesn’t make decisions for me. It leaves a trail of scattered feedback. An AI tool can help organize interview notes into those three columns, but only after I paste in the original words. It shouldn’t guess what the client is thinking from nothing.

Recreate it today: nearly zero cost, start in 10 minutes

Cost: RMB 0. Time: 10 minutes to build the table, plus 5 minutes a day to record entries. Technical barrier: knowing how to copy, paste, and use a document table. First step: create a document titled “Client Feedback Review,” open a new table, and add the columns “Action / Feedback / Next Adjustment.” Then open a real client conversation and fill in your most recent exchange.

If I want to use an existing AI assistant, I can give it that conversation and ask it to extract only “What the client explicitly said” and “One small change to test next time.” I wouldn’t start by asking, “How do I build a personal brand?” That usually produces pages of correct but useless advice.

Use it at the stage I’m at now, not all at once

If I’m just starting, I’d note three feedback instances on paper before building any system. That way, I don’t spend hours organizing feedback when I have no clients whose responses I can observe. Not testing it today is fine too. Once repeated conversations start showing up, I can return to this loop.

If I already have 1–2 reliable clients, I’d review the feedback table once a week, pick the most common preference to change, and archive the replies before and after the adjustment for our next discussion. When a client gives vague feedback, I might ask: “If we changed X this time, would that feel closer to what you mean?”

If I’m scaling up, I’d turn the table into a shared team “feedback rulebook”: what counts as useful feedback, how many variables to change per round, and who confirms the result. AI can handle the summaries, but I wouldn’t let the team mistake a model’s guess for client feedback. Once my own team passes three people, I write the rules down first. Otherwise, adding people just makes the rework move faster.