Originally published in TL;JP, a weekly, fact-checked read on what Japanese AI and tech practitioners are saying on X.
For the week of Sep 29 to Oct 5, 2026, my feed split in two: one small, usable trick for coding agents, and a run of first-person posts about illustration work. Both matter, but not in the same way.
1. A prompt suffix for award-grade front ends
@kajikent says when you want Claude Code or Codex to produce a high-design website in one shot, end the instruction with: aim for quality that could win at awwwards, Webby Awards or FWA, and self-check, iterating until it reaches that bar. That is the whole trick as posted.
https://x.com/kajikent/status/2105095422448746573
Signal — a copy-paste change you can A/B this afternoon.
2. Illustration work, from both sides of the invoice
@charlow_illust, the most-liked post in my set this week at about 13,700 likes, says illustration commissions have nearly stopped since generative AI spread; they call it sour grapes, say they saw it coming, and say work had never dried up before. From the buying side, @EdgeKitune says their company used to commission illustrations, icons and product packaging, but after disputes over contract language forbidding generative AI use, they stopped commissioning that way entirely and now order only through large agencies. @nyaa_toraneko pushes back on the "we're adopting AI, so no designers next month" talk: a company taking responsibility would instead ask how to get its highest-taste people using AI well. All three are individual accounts, not surveys.
https://x.com/charlow_illust/status/2105478334163615846
https://x.com/EdgeKitune/status/2106191479416995899
https://x.com/nyaa_toraneko/status/2106333922942665189
Signal — the contract-clause mechanism is concrete and transferable to any market.
3. The delete-and-remake loop
@guu_guu_fuu took their case to a Bunkacho (Agency for Cultural Affairs) consultation desk, describing a pattern: existing illustrations fed into AI, the output sold on LINE STORE, reported, taken down, then made again. They say one voice may change nothing but hope it adds up. Separately, @aruguchieshi warned an account posting AI variations of their art; the account deleted it at once and apologized, saying it hadn't realized this was unwelcome or that it might be a crime — that characterization is the account's, relayed by @aruguchieshi, who says they don't know how to explain it to people who don't see it as wrong.
https://x.com/guu_guu_fuu/status/2106203378116354278
https://x.com/aruguchieshi/status/2105990108864483698
Signal — enforcement failing on repeat is a product problem, not only a legal one.
4. Noise round
@studio_yebisu posted a personal model tier list for "high-level instruction" work — Opus-5.5 six filled stars, Fable-5.1 five, Astra-6 four filled of five on their own scale, Grok unrated because they don't use it. No criteria given. Noise — taste, not test.
https://x.com/studio_yebisu/status/2105433319802560856
@kensuu says they went looking at crowdsourcing companies' IR data to test whether AI cut freelance work; the post I have is cut off before the answer. Noise — the method sounds right, but the finding isn't in the post.
https://x.com/kensuu/status/2106675090494730737
@tetumemo says "dots" plus working from home may be unbeatable, with bone-conduction Shokz on all day so they can talk constantly, unlimited use right now. The post doesn't say what dots is. Noise — unexplained reference.
https://x.com/tetumemo/status/2105910279104336107
Takeaway: the one thing you can act on today is a prompt change; every claim about disappearing work was one person's experience. Try the first, listen to the second.
Have you written generative-AI terms into your creative contracts yet? Hit reply and tell me what the clause says.
The awwwards suffix, used as a self-grading loop
@kajikent's addition does two things at once: it names external quality bars (awwwards, Webby, FWA) and instructs the agent to keep checking its own output against them until it clears. The second half is the part Western teams usually leave out — most prompts stop at "make it beautiful." Try it exactly as posted on a one-shot page build, and keep the self-check clause when you move it into a project-level instruction file. The post doesn't claim measured win rates, so treat it as a technique one practitioner finds effective, not a benchmark.
https://x.com/kajikent/status/2105095422448746573
What Japanese teams are doing differently
Two things stood out. First, the contract layer is where this is actually being fought. @EdgeKitune's company didn't argue about output quality; it hit a clause about whether generative AI could be used, and responded by routing all commissions through large agencies instead of individuals. If you commission creative work, write your AI-use terms now and decide deliberately whether you can still work with individual creators — otherwise procurement will make that call for you by default. Second, @nyaa_toraneko's framing is a genuinely different staffing question: not "how many designers can AI replace," but "how do we get the people who have been winning this work on taste to use AI well." That's a training budget question, and it's cheaper than rebuilding a design function.
https://x.com/EdgeKitune/status/2106191479416995899
https://x.com/nyaa_toraneko/status/2106333922942665189
Signal — both are decisions you can make this quarter.
A workload reality check for AI-built products
@wiz_mnb has been building a game with AI for nearly three months and doesn't believe mid-to-long-form games will flood the world soon: a genuinely company-grade game isn't a volume of work you finish as a hobby. One builder's view from inside the work, with no data attached — but it's a useful counterweight when someone shows you a demo built in a weekend and extrapolates to a shipped product. Budget for the long tail: content, balance, polish.
https://x.com/wiz_mnb/status/2105987914815664252
Signal — grounded in three months of the poster's own build, not speculation.
Correction, Oct 6: an earlier version of this issue led with token-speed measurements from a post that turned out not to be written by a Japanese practitioner. That item has been removed and the headline changed.
How this is made: I'm based in Japan. AI tools help me collect and translate Japanese posts; every item is checked against the original post before publishing, and each claim links to its source. If this was useful, the weekly issue lands in your inbox at TL;JP.
Top comments (0)