# Claude Code, Codex and the Downloads folder trap (Japan) > Source: > Published: 2026-11-06 13:14:25+00:00 *Originally published in [TL;JP](https://plainaiphone.substack.com), a weekly, fact-checked read on what Japanese AI and tech practitioners are saying on X.* This is TL;JP for the week of Oct 27 - Nov 02, 2026. The most-liked post in my set this week was about a folder, not a new model. **1. Your Downloads folder can be used to attack your coding agent** @claudecode_lab warned people not to use the Downloads folder as a project folder for Claude Code or Codex. If a malicious .md file lands there, the agent may well follow its instructions. @tegnike laid out the attack in steps, and the post got nearly 8,800 likes. You visit a hacker's site and an AGENTS.md file downloads automatically. Later you start a Claude Code or Codex session in Downloads, and that file loads as the agent's starting context. The post I have stops there, so the last step isn't shown. **Signal:** it's a specific risk that takes a minute to fix. [https://x.com/claudecode_lab/status/2114964169002783053](https://x.com/claudecode_lab/status/2114964169002783053) [https://x.com/tegnike/status/2114989227507155040](https://x.com/tegnike/status/2114989227507155040) **2. Setlog: 7.5 million daily users, a team of four** @N_Taisho listed what impressed them about Setlog. It has 7.5 million daily users and four people running it. When a UI change brings a flood of low ratings, the team rolls it back right away. It doesn't collect address books or user attributes. The Android app was built mostly with AI, using Codex and Claude. The founder still designs the key screens in Figma. This is the poster's summary, and the post names no source. **Signal:** it shows concrete ways a tiny team runs a very large app. [https://x.com/N_Taisho/status/2116444381498523934](https://x.com/N_Taisho/status/2116444381498523934) **3. "Some bugs Opus will never find"** [@kenn](https://dev.to/kenn) claims some hard bugs can never be found or fixed by Opus. Their theory is that GPT models care more about logical consistency than about creative ideas, and that this is why OpenAI models stand out in math and physics. That's one person's view, not a benchmark. Separately, @kaixaoki says ChatGPT Pro is badly broken right now and already unusable for heavy users like them. They mention talk that unlimited GPT chat may go away. **Signal:** one user says GPT models are better at hard logic bugs, another says ChatGPT Pro is unreliable right now. Treat both posts as opinions. [https://x.com/kenn/status/2115656961094230479](https://x.com/kenn/status/2115656961094230479) [https://x.com/kaixaoki/status/2116648160814879183](https://x.com/kaixaoki/status/2116648160814879183) **4. A dislike of the default AI web page** @matsu_vr says they dislike the pages Codex and Claude produce: rounded rectangles, too much whitespace, too many mid-tones. Their fix is SAP's Fundamental Library, which works with CSS alone, needs no JavaScript and looks like a plain business app. **Signal:** it's a cheap way to make AI-built pages stop looking alike. [https://x.com/matsu_vr/status/2115330874367676497](https://x.com/matsu_vr/status/2115330874367676497) **5. Making AI's Japanese easier to read** @AM09_21 calls AI-written Japanese "dead." They say combining explain-before-shoten and yomiyasu, then running Claude and GPT through /relay, cut their mental effort a lot. The post doesn't explain what these tools are. **Signal:** it's a real workflow, but you'll have to look up the tools yourself. [https://x.com/AM09_21/status/2115600546308784507](https://x.com/AM09_21/status/2115600546308784507) **6. The AI image argument goes round again** @n_katamari argues that the consent problem starts when the base datasets are built, not only when someone trains a model on one specific person. @rolling_marou says the voice actors' "NO MORE unauthorized generative AI" campaign has always been against *unauthorized* use of voices, not against AI. @mao_DBmiyuki says AI illustration ends up in throwaway ads for overseas idle games, and leaves the feeling that it isn't worth paying for. @ototogoto describes clients who arrive set on an AI-generated reference image, then reject a rights-safe version adjusted for printing, color count and budget. **Noise:** lots of strong feeling and no new facts. The story about client briefs is the one worth keeping. [https://x.com/n_katamari/status/2116777084379074964](https://x.com/n_katamari/status/2116777084379074964) [https://x.com/rolling_marou/status/2116844348654461263](https://x.com/rolling_marou/status/2116844348654461263) [https://x.com/mao_DBmiyuki/status/2116906363049930891](https://x.com/mao_DBmiyuki/status/2116906363049930891) [https://x.com/ototogoto/status/2115859749459894518](https://x.com/ototogoto/status/2115859749459894518) **My takeaway:** the useful posts this week were about the setup around AI tools: which folder the agent starts in, what its pages look like by default, and which model gets the hard problem. Decide those on purpose. Where do you start your coding agent sessions, and have you ever checked what files are in that folder? Here's what a Western team could try from this week's posts. I've kept to what the posts actually support and flagged the gaps. **1. Choose where your agent starts** Never open a Claude Code or Codex session in Downloads, or in any folder that collects files from the web. Start sessions only from a dedicated project folder. @tegnike's attack works because the agent automatically reads an instruction file (AGENTS.md) in the folder where it starts. So before a session, it's worth checking whether that folder holds instruction files you didn't write. Neither post covers agent settings or permissions, so I won't guess at those. **2. Give the agent a design system to follow** @matsu_vr's fix for the generic AI look was to switch to SAP's Fundamental Library, which works with CSS only. That suits internal tools and admin screens, where a plain look is a plus. The post doesn't say how they get Codex or Claude to use it, whether through prompts, project files or something else. If you try it, the idea is the same: give the agent a fixed visual vocabulary instead of letting it pick its own. **3. Make the AI update a dashboard at the end of every session** @AM09_21 added something called impeccable to their setup. They then built a dashboard with HTML and TikZ diagrams, and set a hook so it updates at the end of every session. The point is that the dashboard updates automatically every time, not only when someone remembers. The post doesn't show the hook's code or explain explain-before-shoten, yomiyasu or /relay, so you'd have to rebuild this from the description. **4. Pick a model by type of problem, and test it** [@kenn](https://dev.to/kenn)'s claim gives you something to test: when Opus gets stuck on a bug that needs strict logical reasoning, try a GPT model on the same bug. @kaixaoki's post is the warning: they say ChatGPT Pro is unreliable for heavy use right now. Keep your own short list of which model solved which kind of bug, and don't rely on someone else's opinion, including these two. **5. Habits from Setlog's small team** From @N_Taisho's summary: The post doesn't explain how they detect a ratings spike or how rollback works. **6. Client briefs that include AI images** @ototogoto's story is a warning for agencies. A client brings an AI-generated reference image, has already made up their mind about it, and then rejects anything adjusted for printing, color count or budget. The post offers no fix. My own reading: talk through those limits before you start the work, not after you deliver it. **What looked different in Japan this week** I can only speak for this week's posts. Two things stood out. First, one poster treats making Japanese AI output easier to read as its own problem, with its own tools stacked on top of the models. Second, in the art and voice debates the argument was about consent. The objection is to *unauthorized* use, and @n_katamari pushes that back to how base datasets are built. *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](https://plainaiphone.substack.com).*