X Chinese Topic Brief¶
Translation
English translation of the Chinese original.
Today's Priority Picks¶
-
Turn old computers and old phones into Agent test machines so the main machine stops losing the mouse (corresponds to candidate 1). Worth doing now: this post from Tw93 has 540 likes and 31,700 views and comes with a complete long-form write-up that breaks the concrete annoyance of "agents grabbing the mouse" into three things: dividing labor across devices, permission interfaces, and CI queueing. The best additional angle: when the Chinese-language circle discusses multiple Agents it easily lands on "just buy a few more machines", but what he actually does is give each class of task a screen that will not be grabbed — mouse-moving UI and end-to-end tests migrate to the old machines, while the main machine only handles writing code, signing and notarizing, and shipping releases; one task uses only one system at a time, so an 8 GB test machine does not get dragged to death by a virtual machine; he uses the macOS accessibility interface to read window elements as text and then click, using screenshots only as a fallback. Format suggestion: write a standalone original post, or quote this one and add the applicable conditions: this is one person's device list, not a universal recipe. Timeliness and verification: within 24 hours. The long-form content follows the original post's note; the two tools he wrote himself, ax and input, were not actually tested by this report; "CI costs over USD 70 a month" and "self-hosted has run nearly 180 jobs" are his own bill figures, so do not write them up as an industry benchmark.
-
Slotbound sold 50,000 copies in two days: first check whose numbers these are and what Steam actually disclosed (corresponds to candidate 2). Worth doing now: the Chinese-language circle already has two versions of the numbers and one amplified development story. The developer's own original post says: the first independently developed game, Slotbound, took in more than USD 356,000 in total revenue within 48 hours of launching on Steam, sold 50,000 copies, and peaked at 9,377 concurrent players. These are figures published by the person involved, not an official Steam report, and not a re-test by this report. The best additional angle: align three things before writing — who said the numbers, what the store page actually discloses, and what the positive review rate is now. What is verifiable on the Steam store page: released 2026-11-12, developer and publisher Optima Arch, an About page that describes itself as a single-developer roguelite auto-battler in which a 3x3 slot machine summons units; at the time of crawling the user score was Mostly Positive, with 79% of 868 reviews positive, matching the 79% in the Chinese post, but scores change, so state that this is the page as of this crawl. The AI-generated content disclosure on the same page says that some visual assets and promotional materials were completed with the assistance of generative AI tools and were reviewed and modified by the developer before being integrated into the art style; it does not say "all code generated by AI". Format suggestion: write a standalone original post. Timeliness and verification: representative post 24–48 hours, Chinese second-hand posts within 24 hours. "One year, zero experience, all the code written by AI" belongs to Chinese second-hand retelling and some media reports (one report says "11 months ago had no game development experience at all"), and is not equivalent to the scope Steam discloses — mark it Pending verification, do not write it as established fact.
-
$20 plan comparison chart is popular, but "billions of tokens" does not match official statements; Fable's limited-time inclusion ends this afternoon (corresponds to candidate 3). Worth doing now: this comparison chart with 517 likes and 73,500 views, written in a joking tone, describes the four USD 20 tiers as "3 billion tokens", "870 million tokens", "million-token context plus 5 TB of cloud storage", and "unlimited only at USD 30". Of these, the big token numbers for Claude and ChatGPT have no official source: Anthropic's Pro plan description and usage and length limits say that limits are shared on a weekly basis, that usage varies with message length, conversation complexity, files, model and features, that it can be viewed under Usage in Settings, and do not publish a fixed token count or a fixed number of messages. On Google's side, official materials confirm AI Pro at about USD 19.99 per month including 5 TB of cloud storage, and state expanded use of Gemini 4 Argon in the Gemini app, Antigravity and AI Studio, but the same list of Pro benefits does not say it includes Gemini 4 Carbon. The one with a hard deadline is Fable: Anthropic's Fable 5.2 launch offer description states that Pro plans and Team standard seats, from November 9 through 2026-11-16 23:59 PT, which is 2026-11-17 15:59 Shanghai time, may use up to 50% of their weekly limit on all Fable models, and also receive one Fable quota reset, with both expiring at the same moment; after the 50% is used up or after that week ends, Fable moves to usage credits, and unused 50% does not roll over. At the time this report was written the window had not yet closed, but it ends this afternoon. Format suggestion: write a standalone original post: first what Pro users still have and what happens after expiry, then point out that the big numbers in the chart that are most easily mis-transmitted have no source. Timeliness and verification: within 24 hours. Google's 5 TB and the Fable time window are Verified; the various "billions of tokens" claims and "unlimited Grok" are Pending verification, and the real usage of each plan was not measured.
Candidate Topics¶
There are 10 entries in this issue, closed out by creative value rather than padded to 20. Items 1 to 3 only add information and verification status not expanded in Today's Priority Picks; for specific angles see above.
1. Old computers and old phones converted into Agent test machines: dividing labor matters more than stacking up devices¶
Original post and heat: Tw93 @HiTw93 original post, 2026-11-16 17:53 (Asia/Shanghai), within 24 hours, 540 likes, 31,700 views, 55 replies, 35 reposts, 2 quotes, 496 bookmarks.
Content and focus: the main machine is a black Mac M5 Pro bought this year, 48 GB of memory, 1 TB; writing code, signing and notarizing, and shipping releases all happen on it, and it is also the only machine he uses; in the long-form post he says he regrets not buying 2 TB. The test machine is an M2 Pro that had been gathering dust for months, 32 GB / 512 GB; in October, while testing the Mole software, because an Agent on the main machine was grabbing the mouse and focus, he moved all tests that need mouse control over to it. An iPhone SE3 is plugged into the test machine and operated by AI using iPhone Use and mirroring, running iOS tests for 妙言 and phone tests for a new product; the Agent uses WebDriverAgent to read the screen as elements and then clicks by label; on the same M2, Parallels runs a Windows 11 ARM virtual machine to test Mole for Windows ARM64. There is also an 8 GB Windows ThinkPad X1, used to test screen scaling, battery, SmartScreen and Defender on real x64 hardware and for Windows release acceptance, and inside it an Omarchy Linux virtual machine is nested to test Pake's behavior on Wayland. There is also a NAS mini host running Debian 12 containers, x86_64, 8 cores, 31 GB, responsible for the Linux command line, differences in GNU tools, and Pake's deb/rpm packaging. The test machines all sit in a cupboard in the spare bedroom, plugged in and ventilated, screens at their dimmest and never sleeping, and are restarted together every half month.
Why it is worth doing / possible angles: it can be written as a list of "which jobs must leave the main machine": UI tests that grab the mouse and focus, acceptance that needs a real device or real hardware, and builds that compete with CI for the same screen. His way of handling the conflict between CI and Agents is also concrete — a "working" marker file plus a noclobber-based "CI busy" marker, so that only one can grab it at any moment, with retries and timeouts; Agents also ask each other when the screen will be released. The closing line about priorities, "health > happiness in life > happiness in coding", can be quoted, but there is no need to elevate it into a methodology.
Timeliness and verification: within 24 hours. The device list and tool practices come from the original post's note; ax (reading window elements as text through the accessibility interface) and input (sending mouse and keyboard events with CGEvent) were not tested; the money saved by the self-hosted runner is his own bill, and the point that macOS runners for private repositories are billed at ten times the minutes should be checked against GitHub's billing page, which this report did not verify item by item. Some people in the replies echo "don't let several machines edit the same copy of the code" and "Agents will hog the screen", but there are only 2 quotes, so do not write it up as a heavily quoted discussion.
2. Slotbound: developer says 50,000 copies in 48 hours, Chinese second-hand retelling pushed it to USD 2 million¶
Original post and heat: the person involved, @OptimaArchGames original post, 2026-11-15 12:12 (Asia/Shanghai), 24–48 hours, 1,188 likes, 181,000 views. On the Chinese side, the earliest to give the full game name was @BeyondTheBoxCN original post, 2026-11-16 07:00, 24–48 hours, 203 likes, 20,500 views, 18 replies, 22 reposts, 179 bookmarks, 1 quote; the same author's follow-up post, 2026-11-16 20:00, within 24 hours, 53 likes, 4,159 views, changed the concurrent figure to 9,377, and said the developer can barely program and that all the code in the demo stage was generated by AI, using LÖVE2D with Lua, together with Cursor, Claude and ChatGPT. Another more widely spread post, @threeaus original post, 2026-11-16 11:11, within 24 hours, 145 likes, 33,500 views, 6 replies, 11 reposts, 166 bookmarks, 0 quotes, wrote USD 356,000 and 79% positive reviews, and speculated that over the whole lifetime it might reach USD 2 million.
Content and focus: the USD 2 million is the reteller's speculation, not an official figure. The same author's earlier post saying "15,000 concurrent" is wrong; the developer's own figure of 9,377 should be treated as authoritative. What can currently be verified on the Steam store page is the release date, the developer, the gameplay self-description, the 79% positive rate at the time of crawling, and a passage of AI disclosure about visual assets and promotional materials; "all the code generated by AI" is not within the scope of what the store page discloses. In passing, the policy background can be mentioned: 阑夕 @foxshuo original post, 2026-11-16 19:45, within 24 hours, 35 likes, 18,900 views, about Steam and Epic arguing over whether disclosure of AI use in games should be mandatory. Steam's current practice focuses on AI-generated content in the finished product that players consume; the Steamworks content survey and later clarifications point to the images, sounds, story, dialogue and so on that players will encounter, not to efficiency tools used during development, including code assistants; Epic's Tim Sweeney publicly criticized mandatory disclosure as turning into a stigmatizing label.
Why it is worth doing / possible angles: when writing, keep "what is disclosed is imagery, audio and text" separate from "AI was used to write code", otherwise it is very easy to add in what the store page does not say. Nor should one infer from this single case that "AI can already mass-produce hits".
Timeliness and verification: the representative posts span two time bands. The sales figures are attributed, as a source claim, to the developer's original post on X, and were not cross-validated against an official Steam report; the positive review rate and release date are Verified based on the store page crawled for this report; "one year, zero experience, all code by AI" is Pending verification.
3. The USD 20 tier comparison chart: Google's 5 TB and Fable's deadline are real, the big token numbers are not¶
Original post and heat: @likka_ailive original post, 2026-11-16 18:49 (Asia/Shanghai), within 24 hours, 517 likes, 73,500 views, 61 replies, 7 reposts, 2 quotes, 131 bookmarks. On the same theme there is another post with higher engagement, @JamesDa55049936 original post, 2026-11-15 12:40, 24–48 hours, 316 likes, 143,000 views, comparing ChatGPT Pro at USD 100 with Google AI Ultra 5x at about USD 100.
Content and focus: the comparison chart writes Claude Pro as "3 billion tokens plus a trial voucher", ChatGPT Plus as "870 million tokens", Google AI Pro as "million-token context plus 5 TB for free", and Grok Premium as "unlimited only at USD 30". Of these, the Claude side clearly does not match the official description: Pro is a weekly shared quota, and the official documentation does not publish a fixed token count. The "trial voucher" and Fable's limited-time inclusion are also not the same thing; the latter means up to 50% of the plan's weekly quota can be used on all Fable models, with one quota reset attached, ending at 15:59 Shanghai time today. Grok's "unlimited only at USD 30" is likewise the post's claim, and this report did not check xAI's pricing page. The quote area has only two low-engagement comments, so do not write it up as a lively discussion.
Why it is worth doing / possible angles: this chart works well as an opening hook, not as a price sheet. A workable order: first say which items line up (Google's 5 TB, Fable's time window), then which do not (the token counts for Claude and ChatGPT), and finally the near-term action — what Pro users should confirm before this afternoon, and how the quota behaves after expiry.
Timeliness and verification: within 24 hours. The official statements from Anthropic and Google and the Fable launch offer are Verified; the various big token numbers and Grok's unlimited claim are Pending verification. In the JamesDa55049936 post, claims such as "Codex quota equals 10 times that of Plus" are the author's own statements and were not checked line by line against the official pricing pages.
4. Qwen4-Max official version goes live, and the Chinese-language circle's first reaction is not uniform¶
Original post and heat: official @Alibaba_Qwen launch post; Chinese sample @MichaelLee009 original post, 2026-11-16 11:30 (Asia/Shanghai), within 24 hours, 54 likes, 6,099 views, 46 replies, writing that the official version is here, but "I really want to find an angle to praise it, and it is just too hard".
Content and focus: the official specs (2.4 trillion total parameters, 71 billion activated, average task price down by about 60%, and so on) have already been written up in today's daily report, see the Alibaba Cloud press release and the pricing table; this entry does not repeat them. Discussion on the Chinese side centers on "whether there is a perceptible improvement over the preview version": @leo114119, recorded in the daily report, thinks almost no major change is visible, while @wanxubo writes that it "officially ranks second among Chinese models", while also saying the benchmark scores are mediocre and that it did not reach the expected Fable. Off-list, @xiax0603, 2026-11-16 16:42, 10 likes, 2,196 views, made a visual comparison of "云上龙城" with the same prompt and thinks the official version's texture is not as good as the preview version.
Why it is worth doing / possible angles: one can write "first ask which version and which type of task you are comparing", instead of reporting the parameters once again. The visual comparison is a single sample only, suitable as an entry point for a question, not as a conclusion.
Timeliness and verification: within 24 hours. The official specs were already Verified in the daily report; "worse than the preview version" is one individual's comparison and is Pending verification pending more reproductions. xiax0603's engagement is below the main threshold, so when citing it, note that it is a supplementary sample.
5. Hy4 flash free until November 30, and Chinese discussion settles first on speed and thinking time¶
Original post and heat: @leo114119 original post, 2026-11-16 14:59 (Asia/Shanghai), within 24 hours, 17 likes, 2,486 views; @yedabinbinah original post, 2026-11-16 16:58, within 24 hours, 16 likes, 5,198 views.
Content and focus: the former says the domestic and international versions of WorkBuddy were integrated simultaneously, free for a limited time from November 16 to 30 at 23:59, and that after Hy4's free period ends Hy3 will be taken offline in the app. The latter claims a measured speed of about 150 TPS, that thinking takes rather long, and that overall it is faster than DeepSeek 1017, and mentions that OpenCode and WorkBuddy are opened at the same time. The free period matches Tencent's official launch statement, which was checked in the daily report; 150 TPS is a personal measurement, not an official metric.
Why it is worth doing / possible angles: for Chinese users, what actually needs judging is the deadline, which entry point works, and the experience difference of "fast at output, slow at thinking", rather than reporting once more the 224 billion total parameters and 12 billion activated. Both samples have low engagement, which shows that discussion volume on this topic is currently limited, so note the small sample size when writing.
Timeliness and verification: within 24 hours. The free-period deadline matches the official statement; the speed figure is Pending verification and was not re-tested.
6. The Chinese-language circle is asking: how exactly do you manage Agent versions, memory and cross-device use¶
Original post and heat: @yangyi original post, 2026-11-16 14:21 (Asia/Shanghai), within 24 hours, 75 likes, 38,400 views, 41 replies, 6 reposts, 7 quotes, 145 bookmarks.
Content and focus: the body is a single sentence: "How is everyone doing agent version management, memory management, and cross-device sharing right now?" There are 145 bookmarks but only 75 likes, which suggests quite a few people saved it as a checklist of questions, but the original post gives no answers.
Why it is worth doing / possible angles: the question can be split into three layers — how to version code and prompts, where to put memory and state, and how multiple devices avoid editing the same copy at the same time. The multi-device division of labor in candidate 1 can serve as a comparison: Tw93 solved "who uses which screen", but did not give a general versioning and memory scheme in the original post. This task did not systematically read the 41 replies to that post, so do not invent a comment-section consensus.
Timeliness and verification: within 24 hours. Only the question itself is verifiable; the answer part is an open question.
7. Follow-up: 郭宇 (Guo Yu)'s hayabuild — with cache hits, an iOS build reaches TestFlight in about 38 seconds¶
Original post and heat: @turingou original post, 2026-11-16 17:18 (Asia/Shanghai), within 24 hours, 209 likes, 28,900 views, 7 replies, 5 reposts, 2 quotes, 105 bookmarks.
Content and focus: the Chinese brief of 11-15 covered his remote development path — being able to push iOS and cloud projects forward without a local Mac. This post is the increment: after a few days of tuning, hayabuild has an almost 100% cache hit rate, and one iOS project builds and is automatically packaged into TestFlight beta in just 38 seconds; the next step is to open up slots.
Why it is worth doing / possible angles: write only the increment, do not retell remote development. What can be written is "why the cache hit rate determines waiting time more than machine specs do", and which isolation and quota arrangements would need to be added before such a service is opened to others.
Timeliness and verification: within 24 hours; follow-up. Both the 38 seconds and the near-100% figure are the author's own statements about his own service, not a third-party benchmark, and the service has not yet been opened for public use.
8. Follow-up: DeepSeek Harness's Computer Use is already available as an experimental plugin¶
Original post and heat: @xiaomovps original post, 2026-11-16 16:02 (Asia/Shanghai), within 24 hours, 72 likes, 13,000 views.
Content and focus: he says the official team turned Computer Use into an experimental plugin shipped with DSH, added at the end of October, which he had not paid much attention to at the time; the macOS path is Plugins, More, Experimental Plugins, and after authorization it can observe the desktop and operate applications, still experimental, with the Windows version still under development. The post also mentions the SubAgent, AgentTeam and Git Worktree review in the 0.2.2 series, and that v0.2.2-rc.3 can import DeepSeek App and web history conversations; these are the post's second-hand paraphrase of the changelog.
Why it is worth doing / possible angles: the Chinese brief of 11-14 wrote about DeepSeek Harness as a research prototype; the point this time is "it installs, it can be authorized, it can see the screen", and where the boundaries of an experimental plugin lie. Do not write "added at the end of October" as if it were a new release today.
Timeliness and verification: within 24 hours; follow-up. computer-use.md and the v0.2.2-alpha.2 release notes verify that the capability exists as an experimental optional bundle, that the relevant commit is dated 2026-10-29, and the content of the macOS permission guidance. Desktop operation was not actually tested; nor should importing history conversations be written up as a way to bypass quota limits.
9. One Code: recreating the Claude Code workflow on Pi, with the model swappable¶
Original post and heat: @xiaomovps original post, 2026-11-16 21:12 (Asia/Shanghai), within 24 hours, 77 likes, 7,772 views.
Content and focus: the project is named One Code; it keeps the Pi core and uses a large number of Extensions to rework Plan Mode, Auto Mode, Subagent, Git Worktree and permissions, and it supports reading CLAUDE.md, Skills, Hooks and MCP, without forcing the use of Claude models. The install command is pi install npm:one-code-extension. The author himself also wrote: it is not yet one hundred percent Claude Code, and some cloud capabilities and system-level sandboxing are not implemented.
Why it is worth doing / possible angles: for people who already use Pi but want to keep their Claude Code habits and are unwilling to be locked into one vendor's model, this is a concrete path worth trying. When writing, include both caveats: it is not a complete replacement, and third-party extensions execute code.
Timeliness and verification: within 24 hours. The pi.dev package page shows one-code-extension version 0.8.0, published 2026-11-15, the install command is the same as in the post, the license is MIT, the author is isuruwijesiri, the page states that it can be installed on an existing pi installation, and it carries a warning that third-party packages can execute code and that the source should be reviewed before installing; the repository is IsuruMaduranga/one-code. The star count could not be retrieved reliably, so it is not given. Not actually tested; engagement is ordinary, so it should not be written up as a hit.
10. Follow-up: 宝玉 (Baoyu) says to use Sonnet for scheduling in the main session and hand heavy work to Thread¶
Original post and heat: 宝玉 @dotey original post, 2026-11-16 10:59 (Asia/Shanghai), within 24 hours, 108 likes, 23,100 views, 146 replies, 12 reposts, 5 quotes, 114 bookmarks.
Content and focus: following on from the 11-15 Chinese brief's topic of "how should a person manage a group of Agents", this post gives a division-of-labor suggestion: once the main session plus Thread mode is well optimized, token consumption is acceptable; the main session does not need a particularly strong model, something at the Sonnet 5.5 level is enough, and its job is to identify, route and schedule, like a company's front desk, with many concrete tasks handed to Thread.
Why it is worth doing / possible angles: it can be written as a concrete practice of "the scheduling layer does not need to be the most expensive", and put together with the versioning and memory questions in candidate 6 — scheduling solves allocation, but does not automatically solve where state lives. This is experience sharing, not an official architecture document; this task did not systematically read the 146 replies, so do not invent a comment-section consensus.
Timeliness and verification: within 24 hours; follow-up. Only this one scheduling suggestion is new, and the judgment about token consumption is personal perception.