待翻譯：How to get better every week with AI

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Adam Faik Aug 03, 2026 A few years ago, at a previous company, I sat through a leadership training I’ve mostly forgotten. One idea survived. The most senior product leader in the room told us to end every week the same way: write yourself three or four honest lines about how the week went, and what you’d do differently on Monday. His pitch was simple math, borrowed from the idea James Clear made famous in Atomic Habits: get 1% better every week, and compounding does the rest. I’ve written that little Friday retro almost every week since, and it works. For years, though, the habit had a flaw I couldn’t see. The reviewer was me. My memory of the week is a friendly narrator: it remembers the demo that landed, and it quietly forgets the question I dodged in a tense meeting. Every Friday I graded my own homework, and every Friday I passed. A retro written from memory is a review of the week you’d like to have had. Better than nothing, but it plateaus fast. Then one afternoon I pasted the verbatim transcript of a user interview I’d just run into Claude. Not to summarize it (I do that all the time). I asked it to review my interviewing: how I phrased questions, what I missed, how much I talked. It found three leading questions in my first ten minutes and quoted them back word for word. Nobody had reviewed my work at that level of detail in years. Everyone I know uses AI to do the work: draft the spec, summarize the research, write the update. Almost nobody uses it to review how they work. The first use saves you hours this week. The second one compounds, because it makes every next interview, meeting, and decision slightly better than the last. That’s the 1% from the training, and AI finally makes it cheap to collect. The whole practice fits in one picture. The gap. I’ll show you why nobody reviews your actual craft anymore, and why the feedback you do get says more about the giver than about you. The loops. The three I run (interviews, meetings, the Friday retro) with the exact prompts I use, plus six more worth trying, from your forecasts to your 360 packet. The honesty problem. Here’s how to stop the model from flattering you, because by default it will. The team angle. You’ll build a skill that asks your PMs your own 1:1 questions before they ever reach you. By Friday you can have run your first loop on a real artifact from your own week. You get the kind of specific, evidence-quoting feedback nobody has given you since you got senior, and you can get it every single week. The first loop needs nothing new: any LLM chat works (Claude, ChatGPT, Gemini), and the raw material is the transcripts your meeting tools already produce (Zoom AI Companion, Gemini in Google Meet, Granola, or a voice memo on your phone). And if you lead PMs, you’ll leave with two skills worth an afternoon each: one that reviews your week on a schedule, one that preps your team for 1:1s. Your reviewer is ready. Let’s put it to work. Nobody reviews your work anymore Think about the last time someone watched you work and told you something true about it. Not your outcomes: your craft. The way you ran the interview, the way you handled pushback in the room. If you lead product teams, the honest answer is probably “years ago.” Your manager sees headlines and results, not how you got them. Your team won’t critique you upward. The more senior you get, the less anyone reviews how you actually work. What’s left is the official cycle: twice a year, compressed into themes like “communicate more.” And hold even that loosely. In The Feedback Fallacy (Harvard Business Review, 2019), Marcus Buckingham and Ashley Goodall assembled the research on how well humans rate other humans. The headline finding has a name, the idiosyncratic rater effect: more than half of any rating of you reflects the rater, not you, and no amount of training fixes it. In their words, feedback is “more distortion than truth.” Your manager’s read on your “communication” says as much about their definition of communication as about yours, and that stays true for the best managers you’ll ever have. The little feedback that reaches you arrives filtered through someone else’s lens, and the research says the filter is most of the signal. Not a reason to dismiss your manager. A reason not to make two filtered data points a year your only mirror. Performance-heavy fields treat this as a solved problem. Athletes and musicians review tape after every game and every recital. Atul Gawande wrote the canonical essay on this, Personal Best (2011): a surgeon at the top of his field plateaued after eight years, then started improving again when he put a reviewer back in his operating room. The research on expertise agrees. Anders Ericsson’s work on deliberate practice (Peak, 2016) found that experience without feedback doesn’t accumulate into skill; it plateaus. Twenty years of interviews don’t make you better at interviewing if nobody ever shows you what you did. What changed isn’t the theory. It’s the tape. Your meeting tools record by default now: Zoom AI Companion, Gemini in Google Meet, Granola quietly taking notes on your laptop. Verbatim records of you working have been piling up in your drive for a couple of years. We skim the summaries for action items and move on. The evidence of how you work is exhaust of a normal week, and almost nobody reads it in the review direction. One precision before you touch that pile. The AI notes your tools generate are summaries, and a summary is already the friendly narrator’s cut, machine-made this time. The loop needs the verbatim transcript. In Google Meet, that’s the separate “Transcripts” option, not the Gemini notes doc. In Zoom, it’s the transcript file attached to the recording. Granola keeps the full conversation one click behind its polished notes. Feed the model what was actually said, not what the notes decided to remember. And the reviewer for all that unread tape? You already have it. You’ve just been pointing it the other way. You already run AI as a production engine. It drafts the spec, summarizes the research, writes the launch note. Point the same model at a transcript of how you worked, and it becomes a feedback engine. Same tool, opposite direction. In that direction it has qualities no human reviewer can offer: the patience to re-read your entire week verbatim, no stake in your ego, and availability at 6 pm on a Friday. The only reviewer with time to watch all your tape has been sitting in your browser tab all along. The direction flip is the whole method. Here’s what it looks like on a real week. Start with my three loops I run three loops on my own weeks, and they share one shape. Take a verbatim artifact your week already produced. Ask the model to review how you worked, not what was decided. Extract one change for next time. Write that change down. The artifact does the honesty, the model does the patience, and the written line makes it stick. Here’s the shape before the details. Three isn’t a magic number, and my three aren’t the menu. They’re worked examples, mapped to where my own weeks leave a record: user interviews, hard meetings, and the week itself. Treat them as illustrations of the shape, not the full list; your calendar will suggest loops mine can’t. A catalog of other places to point the same shape comes right after these three. Loop 1: Review your interviews After a user interview, I copy the full transcript into the LLM and ask it to review the interviewer. Not the user’s answers, not a summary of insights: me. The phrasing of my questions, the follow-ups I didn’t ask, how much of the conversation was my own voice. The distinction matters because “summarize this interview” is production work, and every PM already does it. ”Review how I ran this interview” is feedback work, and it’s the version almost nobody asks for. Here’s the prompt I use, lightly cleaned up. Yours will look different; the load-bearing parts are the named rubric and the quoted evidence. Here is the verbatim transcript of a user interview I ran today. Review the interviewer (me), not the user. Grade me against The Mom Test’s three rules: talk about their life, not my idea; ask about specifics in the past, not hypotheticals about the future; talk less, listen more. For every finding, quote the exact line from the transcript. List every question I asked, label each one open or closed, then tally the list. End with the one change that would most improve my next interview. The rubric is The Mom Test, Rob Fitzpatrick’s 2013 book on customer conversations, and spelling its rules out inside the prompt matters. The critique works even if you (or the model) never opened the book, and you learn the rubric by reading your own violations of it. And the leading questions were only the loudest finding. The same session flagged the moment a user mentioned a workaround and I moved on instead of digging. It listed my questions, and over half were closed. None of that was in my notes, because my notes were about the user. The critique is the start, not the verdict. I discuss it, and the discussion is where the learning happens. The rule I follow: seek evidence, don’t argue. “Show me the line” is a safe question. “Rewrite that question the way I should have asked it” is a great one. Protesting (”I don’t think that was leading”) is the one move to avoid, because models tend to fold when you push back, and a critique that folds is worthless. Treat the chat like a film session with a coach, not a negotiation over your grade. Interviews are the easiest tape to start with: low stakes, famous rubric, fast wins. The harder tape is the meeting where you had something to lose. Loop 2: Debrief your hard meetings Some meetings matter more than others. The technical review where the architecture pushback derailed your agenda. The leadership meeting where your update ran long and the ask got lost. Your calendar is full of them, and that’s not the failure it’s usually made out to be: meetings are where product work happens. Spending most of your week in meetings isn’t the problem. Letting them end with nothing actionable is. I used to replay the hard ones in my head on the way home, which felt like reflection but was really just rumination. Now I feed the transcript (or my detailed notes, when there’s no recording) to the LLM and ask for a debrief of my own performance. How I presented, where I lost the room, how I handled the pushback, what to change next time. The hours were already spent; the debrief is how they pay something back. This move has a precedent you already practice at the team level. Product teams close every sprint with a retrospective: what went well, what didn’t, what changes next time, run by the team, for the team. Norman Kerth, who wrote the practice’s handbook (Project Retrospectives, 2001), even gave it a safety rule, the Prime Directive: whatever we discover, everyone did the best they could with what they knew at the time. Yet the ritual usually stops at the team boundary, and nobody runs one on their own performance in a single meeting. A meeting debrief is a one-person retrospective, and AI makes it cheap enough to run after every meeting that matters. The prompt is shorter than the interview one, because there’s no famous rubric for meetings. I anchor it on my goal instead. Here is the transcript of a technical review I ran today. My goal was alignment on shipping the migration in two releases instead of one. Review my performance, not the decisions: how I opened, how I responded to pushback, where I talked past someone, where I lost the thread. Quote the line for every finding. Then give me the strongest case that this meeting went badly for me. End with what I should do differently in my next meeting with this group. The debrief pays out twice. Once right after the meeting, when it shows you the exact moment you started answering a different ques [truncated for AI cost control]