Vol.067 — AI Native, the Japan Edition (Day 3): If All You Ever Say Is Yes, the AI Never Grows
Date: 2026-08-20 | Newsletter
Key Summary
Vol.067 is Day 3 of four with the CEO and a colleague at a Japanese cosmetics maker. It opens with a report from the client: she had hit her ChatGPT usage limit and moved from the USD 20 plan to the USD 100 plan — five times the cost — because she had spent the week working through the online shop’s SEO homework with the AI Webmaster. Zenta reframes that number rather than apologising for it. The usage rose because work that used to go to an outside web agency is now being run in-house; USD 100 is not a hobby expense, it is a new outsourcing cost, and against what the same work used to cost it runs at roughly one-twentieth. Two weeks after “we outsource because nobody here can write HTML,” she is in a position to review that arrangement whenever she likes. Whether to switch is the client’s call; that the choice exists at all is, he writes, the most concrete result of the training so far. The homework was done properly — product copy, search-critical fields, images — and the happiest stretch of the day was re-running the health check so she could see for herself what had improved. Before the training she could not have named the problem; now she can find it, fix it, and verify it. Once that loop turns, he says, you are fine. The substance of Day 3 was not a new feature but how to raise an AI. Watching the pair work, Zenta noticed that almost every proposal from the AI Webmaster was met with “yes.” The work gets done and the output is not bad — but this is no different from a human subordinate: someone who only ever says yes gives you nothing to adjust against. The best operators say clearly what they like and dislike, and give the reason. Putting the reason into words makes you check your own standard, so the exchange doubles as thinking practice — and the place to keep those reasons is the memory file and the preference file. His phrase for the day: dye your AI staff your own colour. Three unglamorous habits followed. Start each piece of work in a fresh conversation inside the project, so rules agreed mid-stream actually take effect. When you notice something, give the reason and say where to record it — standing rules to the instruction file, company facts to memory, this week’s progress to the status file — because skipping that means explaining the same thing every time. And in design work the preference file matters most, especially the reasons for rejections: “this is wrong, the whitespace is too tight” carries far more information than “this is fine,” and keeping it visibly cuts the back-and-forth. Asked whether the AI Webmaster could also design a printed brochure, he said yes — the time spent aligning and the trust built become an asset, and this colleague does not change jobs or forget what it learned, so the range you can delegate widens on its own. Asked about weekly and monthly activity reports, he drew a line. The background research can be handed over, and they set it to run automatically on a fixed weekday. Two things cannot. The weekly report itself, because outsourcing it outsources the habit of reflecting. And minutes — especially of sales meetings — because AI can transcribe what was said but cannot decide what mattered, and in Japanese the meaning often sits in what was left unsaid. Widening the delegation and drawing the line, he notes, always come as a pair. Do it simple — your usual work, in your usual words.