X Chinese Hot Topics Brief¶

Date22 October 2026 · ThursdayWindow2026-10-20 08:27 to 2026-10-22 08:27CoverageComplete

Translation

English translation of the Chinese original.

Today's top picks¶

The dsh plugin ecosystem, and "handing phone tasks to DeepSeek" (see item 1). What is new this round is not model capability but the ring of plugins that has grown up around DeepSeek Harness, plus one new way of using it: @xiaomovps listed 5 plugins in one post (plugin market, layered compression for long conversations, adding Bash/PowerShell 7 to Windows, multi-backend search, and file editing plus Git in the sidebar), with 294 bookmarks ahead of 214 likes; @realchendahuang wrote about a plugin that connects a phone to dsh on a computer; @turingou, because Codex quota burns too fast, hands device-operation tasks such as iPhone use and Android use to DeepSeek Flash connected through Magpie. There is also a high-engagement post on the other side: @goan999999 pushes dsh-our-free-model, "no registration, no API Key, use a free upstream channel", with 234 likes, but the upstream source is unclear, so it can only serve as a counterexample. The best supplementary angle is a "minimum viable dsh plugin set" for non-programmers, together with a clear list of what not to install: an unknown free model channel means handing your conversation to a relay you know nothing about. Form: a standalone original post with a "what to install / why / risk" table. Status: ready to draft, plugins need hands-on testing; the claim by @yetone that "dsh's user numbers have fully surpassed pi" comes with no data source and must not go into the conclusion.
"Permanent" caching for logged-in Codex and Work users (see item 2). The factual source is solid: the 10-21 09:10 English post from Codex lead Tibo @thsottiaux, with 2.436 million views and 15,000 likes, says the API can now cache for 12 hours, while caching for logged-in Codex and Work users is kept permanently, that returning to an old conversation or resuming hours later carries no penalty, and that a cache hit is guaranteed. The Chinese reaction splits three ways: @dotey calls it "the most valuable update in days" and speculates that it is written to disk (he labels it as speculation himself); @off_thetarget explains that the cache is tied to logged-in users; @BigAvenu3 directly questions the size of the KV Cache and asks whether this is a technical breakthrough or just a benefit whose cost is controllable. The best supplementary angle is to do the math for ordinary users: what resuming a long conversation actually saves, whether quota lasts longer, and how "permanent" is actually billed — the official word so far is a single sentence. Form: a standalone original post, or a reading of Tibo's original post. Status: ready to draft, billing details and any cap still need verification; the cache durations @Gorden_Sun gives for the Claude client (1 hour), the API (5 minutes) and the DeepSeek API (3 days) are his own claims and unverified.
Four "proactive" personal assistants, and how to choose (see item 3). The daily report already covered the Claude Duet launch, so this only takes the selection angle: @dotey counted four similar products from four companies in just over two months — xAI Grok Bot (August 11), Meta Muse (September 8), OpenAI Dots (September 29), and Anthropic Claude Duet; his other post, the Duet breakdown, lays out the boundaries clearly: research preview, open only to Pro/Max in the US, waitlist required, no separate charge during the preview but shared quota, checks the account every few hours, requires user approval before sending messages, making payments, signing documents or deleting things, and will not trade securities or transfer money even with approval. The most interesting discussion sample is @some_local_ai with "the four big labs are queuing up to stuff me with identical cyber housekeepers, and my little to-do list is not enough to feed them", along with @dingyi criticising Grok Bot for being too text-heavy and having no aesthetic sense. The best supplementary angle is not to compare feature lists but two more practical things: whether your to-do volume is enough to feed a proactive assistant, and how to set authorisation boundaries (which actions must be confirmed every time). Form: a standalone original post. Status: ready to draft; Duet is currently open only to Pro/Max users in the US, so Chinese users cannot use it yet, and the post must say so.

Candidate topics¶

1. The DeepSeek Harness plugin ecosystem, and handing phone tasks to DeepSeek¶

Original posts and engagement:
@xiaomovps｜original post｜10-21 17:35 Shanghai｜within 24 hours｜18,000 views, 214 likes, 294 bookmarks
@realchendahuang｜original post｜10-21 01:11 Shanghai｜24–48 hours｜19,000 views, 187 likes, 277 bookmarks
@turingou｜original post｜10-21 21:54 Shanghai｜within 24 hours｜33,000 views, 139 likes
@LinearUncle｜original post｜10-20 20:15 Shanghai｜24–48 hours｜16,000 views, 76 likes
@yetone｜original post｜10-22 01:15 Shanghai｜within 24 hours｜39,000 views, 189 likes
Counterexample: @goan999999｜original post｜10-21 18:27 Shanghai｜within 24 hours｜20,000 views, 234 likes
Content and what people are focusing on: the 5 plugins xiaomovps lists are dshmarket (plugin market), billion-context (layered compression for long conversations), dsh-ops (adds Bash/PowerShell 7 to Windows, Windows x64 only), dsh-free-search (Bing/DuckDuckGo/SearXNG, multiple backends), and dsh-better-sidebar (file editing, terminal, Git and Diff Review integrated). realchendahuang wrote about a plugin that connects a phone to dsh on a computer. LinearUncle uses Magpie to switch to the opencode go plan's deepseek flash inside dsh with one click. turingou, because Codex quota burns too fast, hands device-operation tasks such as iPhone use and Android use to DeepSeek Flash connected through Magpie. yetone says "dsh's user numbers have already fully surpassed pi" without giving a data source. goan999999 pushes dsh-our-free-model, "no registration, no API Key, use a free upstream channel", with an unclear upstream source.
Why it is worth doing / possible angles: the plugin ecosystem is the only genuinely new element this round, and it is also the part ordinary readers can most easily put to use. You can write a "minimum viable dsh plugin set": install only the three to five that actually solve a problem, and explain the payoff of long-conversation compression, the sidebar and multi-backend search; then write "what not to install", using the high-like free-upstream post as a counterexample to explain why an unclear relay deserves caution.
Timing and verification: mostly within 24 hours, plus two items in the 24–48 hour range. dshmarket is a community plugin: the README of dsh-market GitHub says the catalogue has 4,200+ plugins while the dshmarket.com homepage says 6,254, so the two numbers disagree; the install command is dsh plugin --profile web add dshmarket; the official docs include a plugin-market page, DeepSeek Harness docs. Community plugin quality and security are unaudited and nothing was tested hands-on this round; yetone's user-number comparison has no source and remains to be verified. This is a follow-up: item 7 of the 10-21 brief covered how the same model performs very differently under a different harness; the increment here is the plugin ecosystem and phone control.

2. "Permanent" caching for logged-in Codex and Work users¶

Original posts and engagement:
@thsottiaux｜original post｜10-21 09:10 Shanghai｜within 24 hours｜2.436 million views, 15,000 likes
@off_thetarget｜original post｜10-21 11:54 Shanghai｜within 24 hours｜36,000 views, 61 likes
@dotey｜original post｜10-21 13:17 Shanghai｜within 24 hours｜30,000 views, 87 likes, 82 replies
@BigAvenu3｜original post｜10-21 11:59 Shanghai｜within 24 hours｜8,348 views, 37 likes
@Gorden_Sun｜original post｜10-21 11:36 Shanghai｜within 24 hours｜6,943 views, 48 likes
Content and what people are focusing on: Tibo's post is in English and is labelled "Day 16"; it says the API can now cache for 12 hours, while caching for logged-in Codex and Work users is kept permanently: returning to an old conversation or resuming hours later carries no penalty, and a cache hit is guaranteed; the post cites the 12-hour cache announcement from @OpenAIDevs. "Day 16" continues the "28-day improvement plan" covered in the 10-20 brief. On the Chinese side, dotey calls it "the most valuable update in days" and speculates that it is written to disk (he explicitly labels it as speculation); off_thetarget explains that the cache is tied to logged-in users; BigAvenu3 questions the size of the KV Cache and asks whether this is a technical breakthrough or a benefit whose cost is controllable; Gorden_Sun compares it with a 1-hour cache in the Claude client, 5 minutes for the API and 3 days for the DeepSeek API.
Why it is worth doing / possible angles: for ordinary Codex users the real impact comes down to one question — when you go back to a long conversation, are you no longer "paying again", and does your quota last longer. You can write along that question, then add a note on how "permanent" is billed and whether there is a cap: the official word is a single sentence, so say plainly that the rest is unknown.
Timing and verification: everything within 24 hours. Tibo's post is a statement from an official figure, but no accompanying document explains what "permanent" means exactly; dotey's "written to disk" is speculation; the three cache durations Gorden_Sun gives are his own claims and remain to be verified. Only the text of the English original was read; the quoted context was not. This is a follow-up: the 10-21 daily report covered the API's 12-hour prompt cache; the increment here is permanent caching for logged-in Codex and Work users.

3. Four "proactive" personal assistants, and how to choose¶

Original posts and engagement:
@dotey｜original post｜10-22 05:08 Shanghai｜within 24 hours｜22,000 views, 157 likes, 143 bookmarks
@dotey｜Duet breakdown｜10-22 03:54 Shanghai｜within 24 hours｜30,000 views, 71 likes
@dingyi｜original post｜10-21 22:29 Shanghai｜within 24 hours｜9,595 views, 63 likes
Discussion sample: a reply from @some_local_ai｜reply｜10-22 04:43 Shanghai｜within 24 hours｜21,000 views, 4 likes
Content and what people are focusing on: dotey laid out the four similar products from four companies over just over two months: xAI Grok Bot (August 11), Meta Muse (September 8), OpenAI Dots (September 29), and Anthropic Claude Duet. The Duet details: research preview, open only to Pro/Max in the US, waitlist required, no separate charge during the preview but shared quota, checks the account every few hours; requires user approval before sending messages, making payments, signing documents or deleting things; will not trade securities or transfer money even with approval. The official description is in the Claude Duet introduction. dingyi argues that many apps are turning into personal agents, and criticises Grok Bot for being too text-heavy and having no aesthetic sense; some_local_ai's reply is "the four big labs are queuing up to stuff me with identical cyber housekeepers, and my little to-do list is not enough to feed them".
Why it is worth doing / possible angles: the four feature lists overlap heavily, so what really separates them is two more practical things: whether your to-do volume is enough to feed a proactive assistant (with too little to do, proactivity just becomes interruption), and how to set authorisation boundaries (which actions need confirmation every time, which are never authorised). You can follow Duet's approval list to give readers an authorisation tiering of their own.
Timing and verification: everything within 24 hours. Duet's status (research preview, US Pro/Max only, waitlist required) comes from dotey's breakdown and the official page link is included; Chinese users cannot use it yet, and the post must say so. "Not enough to feed them" is a joking opinion, not a conclusion.

4. A one-night "combo" of security updates: what small site owners should check first¶

Original posts and engagement:
@Vincent_AINotes｜original post｜10-21 23:55 Shanghai｜within 24 hours｜20,000 views, 171 likes
@MaxForAI｜original post｜10-21 21:43 Shanghai｜within 24 hours｜29,000 views, 90 likes
@op7418｜original post｜10-21 23:39 Shanghai｜within 24 hours｜25,000 views, 63 likes
@dotey｜Railway post｜10-22 06:37 Shanghai｜within 24 hours｜6,272 views
@dotey｜TiDB Cloud post｜10-22 06:15 Shanghai｜within 24 hours｜8,649 views
@dotey｜GitLab post｜10-22 04:57 Shanghai｜within 24 hours｜9,003 views
Content and what people are focusing on: Vincent_AINotes' roundup of one night: PHP urgently fixed two critical vulnerabilities, Node.js fixed 10, nginx fixed 5, Go updated certificate trust validation, and the eMudhra CA was blacklisted by browsers. MaxForAI posted a PHP2Shell warning and said it involves "56% of the global internet". op7418 says eMudhra used easily guessable numbers as private keys. dotey posted three in a row: Railway's key-storage system was breached and the official instruction is for users to replace all keys; PingCAP notified TiDB Cloud customers to rotate API keys immediately; GitLab had its third full-score vulnerability in six weeks.
Why it is worth doing / possible angles: give non-programmers who built a small site or tool with AI a "what to check tonight" list: first the PHP version, then whether the hosting platform's keys need rotating, then whether a self-hosted GitLab is patched. The order matters more than the enumeration, because most people close the tab after five security headlines.
Timing and verification: everything within 24 hours. The PHP part was already verified in the 10-21 daily report: PHP officially released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 on October 21. The rest is personal roundup or relay and remains to be verified: the "56% of the internet" claim is unverified, the numbers 10 for Node.js and 5 for nginx were not checked against official announcements, and the Railway, TiDB Cloud and GitLab items are all dotey's relay and need to be checked against official announcements before drafting.

5. Modding a Xiaomi TV launcher with AI: only the apps you use, plus wallpaper rotation¶

Original posts and engagement:
@op7418｜original post｜10-20 23:58 Shanghai｜24–48 hours｜57,000 views, 247 likes, 240 bookmarks
@op7418｜tutorial post｜10-21 17:15 Shanghai｜within 24 hours｜22,000 views, 207 likes
Content and what people are focusing on: op7418, referring to CleanFrame and cleanyourtv, repackaged the guizang TV launcher to show only frequently used apps (PS5, Apple TV connection, Bilibili, screen casting) and added wallpaper rotation; it applies to Xiaomi TVs. The next day he added that "removing ads from TVs with AI is very popular lately" and recorded a detailed tutorial.
Why it is worth doing / possible angles: removing ads from home appliances is one of the most tangible AI uses for ordinary people, more intuitive than any benchmark. You can write it as a follow-along test: from a clean launcher to wallpaper rotation, while spelling out the risks of a third-party launcher — whether system updates overwrite it, warranty, and where the install package comes from.
Timing and verification: one item 24–48 hours, one within 24 hours. Author's own account, not tested this round; replacing the system launcher with a third-party one involves system updates, warranty and install-package provenance risks, so it is worth running it yourself and taking screenshots before drafting. This is a follow-up: the 10-20 English brief covered the trend of stripping redundancy out of devices and following along; the increment here is a Chinese-language Xiaomi TV tutorial.

6. iPhone Duo review embargo lifts: apps without large-screen support still render at phone size¶

Original posts and engagement:
@op7418｜original post｜10-21 20:56 Shanghai｜within 24 hours｜126,000 views, 88 likes
@_FORAB｜original post｜10-21 18:15 Shanghai｜within 24 hours｜51,000 views, 84 likes
@abskoop｜original post｜10-21 14:54 Shanghai｜within 24 hours｜63,000 views, 63 likes
@kenw_2｜original post｜10-21 20:29 Shanghai｜within 24 hours｜12,000 views, 12 likes
@waylybaye｜original post｜10-20 18:47 Shanghai｜24–48 hours｜26,000 views, 103 likes
@yihui_indie｜original post｜10-21 08:02 Shanghai｜24–48 hours｜7,661 views, 24 likes
Content and what people are focusing on: after watching a pile of reviews, op7418 jokes that it is "not adapted at all"; _FORAB says apps without large-screen support still render at ordinary phone size when unfolded, and says Apple mandates adaptation before April next year; kenw_2 relays the MKBHD review: a clear first-generation feel, heavy, thick, all-day battery, poor 4x zoom, and bad app adaptation. On the positive side, abskoop says Alipay is already adapted. From the developer side: waylybaye says the simulator only ships with Settings and Safari, so there is too little design to reference; yihui_indie built a free tool that turns a screenshot or screen recording into an iPhone Duo open-close animation.
Why it is worth doing / possible angles: set adaptation priorities for indie developers: first make the app not look bad in the unfolded state, then talk about new large-screen interactions; along the way, cover what AI can help with (layout rework, previews for two screen sizes). This draws more readers than relaying reviews and fits readers' actual situation better.
Timing and verification: mostly within 24 hours, plus two items in the 24–48 hour range. abskoop says Alipay is already adapted, which is the author's own observation. To be verified: "Apple mandates adaptation before April next year" is _FORAB's claim, with no official document seen; the MKBHD review content is kenw_2's relay and was not checked against the original video.

7. One person plus AI making games: from porting old titles to shipping on Steam¶

Original posts and engagement:
@DashHuang｜original post｜10-21 15:35 Shanghai｜within 24 hours｜35,000 views, 305 likes
@AlchainHust｜original post｜10-21 12:07 Shanghai｜within 24 hours｜32,000 views, 300 likes
@Pluvio9yte｜original post｜10-21 17:08 Shanghai｜within 24 hours｜11,000 views, 143 likes
@onefifthblue｜original post｜10-21 23:56 Shanghai｜within 24 hours｜4,789 views, 68 likes
Content and what people are focusing on: DashHuang made a Vision Pro version of Torchlight 2 with controller, touch and eye-tracking support, and is about to ship a TestFlight build supporting Vision Pro, iOS, Mac and Apple TV at once; he had previously modernised the Mac version. AlchainHust recommends @Nenlyonis: a designer by background with 710,000 followers on Bilibili, who taught himself with AI and independently made a Steam game that drew 100,000+ players in two months after a free launch. Pluvio9yte used Claude Code (Opus 5.5) with ViggleAI's PINOC and a web 3D engine to turn "The Peach Blossom Spring" into a third-person adventure game; the key method was comparing fixed-camera screenshots side by side with reference works, iterating one scene eight times. onefifthblue put together a list of domestic game distribution platforms (TapTap, Haoyou Kuaibao, App Store China, and WeChat/Douyin/Kuaishou/Baidu mini games).
Why it is worth doing / possible angles: "making it" is no longer hard; what is hard is benchmarked iteration and shipping to earn money. Pluvio9yte's eight rounds of comparison are the most replicable method this round and can be written up directly as a tutorial; onefifthblue's platform list covers the last leg of the journey.
Timing and verification: everything within 24 hours. All author accounts; the "710,000 Bilibili followers and 100,000+ players in two months" figures for Nenlyonis are AlchainHust's relay and need verification; the other projects were not independently verified.

8. The full Mole Windows development log: rules can be reused, code cannot¶

Original posts and engagement:
@HiTw93｜original post｜10-21 22:36 Shanghai｜within 24 hours｜12,000 views, 239 likes, 169 bookmarks
Content and what people are focusing on: the second long post. The Windows version uses C# + WinUI 3 + .NET 10; of more than a hundred thousand lines of Swift, less than 20% could be reused outside the system frameworks, so it was written almost from scratch; design tokens and copy in 11 languages are synced from the Mac version by script; SYNC.md is used to reconcile features across the two sides; signing uses a Certum cloud OV certificate (just over 300 US dollars for 3 years), and Microsoft Azure Artifact Signing is open to individuals only in the US and Canada; Windows runs in Parallels on a Mac, with remote compiling and debugging through prlctl exec; 80,000+ lines of code and 4,000+ tests in one month, most written by Opus 5.5; a ThinkPad is used for on-device testing.
Why it is worth doing / possible angles: one sentence works as the headline — "in the AI era what you can reuse is design, copy and rules, not code". For indie developers considering a second platform, this offers a very concrete migration ledger: how much is reusable, how to buy signing, how to test on real hardware, how to write the tests.
Timing and verification: within 24 hours. Entirely the author's own account (the second long post, following the tech-stack selection piece that was item 9 in the 10-21 brief), not independently verified; the certificate price and the scope of the Azure policy are worth checking on the official page before drafting.

9. QuestMobile Q3 AI report: Doubao at 419 million MAU, WorkBuddy leading office agents¶

Original posts and engagement:
@foxshuo｜original post｜10-21 19:26 Shanghai｜within 24 hours｜12,000 views, 33 likes
@geekbb｜original post｜10-21 09:13 Shanghai｜within 24 hours｜21,000 views, 75 likes
@geekbb｜follow-up post｜10-21 15:58 Shanghai｜within 24 hours｜9,670 views, 12 likes
Content and what people are focusing on: foxshuo excerpts the report, saying Doubao is 2.5 times Qwen and 3.5 times DeepSeek, that Kimi cites the most sources, and that Simplified Chinese PC web AI monthly actives are declining. The report's official numbers (September 2026): AI-native apps at 546 million MAU; Doubao 419 million, Qwen 158 million, DeepSeek 123 million; AI office at 142 million MAU, with WorkBuddy at 32.353 million, Baidu Dazi at 9.77 million and Qwen Office at 7.905 million. By the official numbers, Doubao is about 2.65 times Qwen and 3.4 times DeepSeek, so the 2.5x and 3.5x in the post are approximations. geekbb says he uses WorkBuddy because check-ins earn 100 credits, and later says he "probably won't recommend WorkBuddy again"; the reason is not given.
Why it is worth doing / possible angles: Chinese tech circles talk about Claude Code all day, but the mass market uses Doubao and WorkBuddy — that gap is the topic. You can write about how content creators should read it: follow the mass market with tutorials, or keep serving the heavy users in the niche.
Timing and verification: everything within 24 hours. The report page is the QuestMobile report page, and the core MAU figures were checked against it; the claim that "WorkBuddy has more than the 2nd to 10th places combined" could not be verified in the summary, so do not write it; "Kimi cites the most sources" and "Simplified Chinese PC web AI monthly actives are declining" currently rest only on foxshuo's relay. This is a follow-up: item 4 of the 10-21 brief covered ordinary people using Doubao to fix their computers; the increment here is quarterly data.

10. AI replies flood X, and Chinese authors start building their own tools¶

Original posts and engagement:
@dotey｜original post｜10-20 22:36 Shanghai｜24–48 hours｜21,000 views, 69 likes
@vista8｜original post｜10-21 14:50 Shanghai｜within 24 hours｜9,198 views, 81 likes, 76 replies
@tualatrix｜original post｜10-20 11:23 Shanghai｜24–48 hours｜7,741 views, 22 likes
Content and what people are focusing on: dotey says AI replies are not limited to Chinese-language circles — all of X is full of them — and hopes the platform will address it; he tests bots with prompt-injection style replies. vista8 open-sourced a browser extension that uses the Jev model to identify AI comments and lets you block manually, plus features such as @ optimisation and bookmark export. tualatrix uses Grok Bot to scan spam replies, giving evidence and options, with human review before writing to an audit sheet.
Why it is worth doing / possible angles: protecting your own comment section is moving from complaining to building tools. You can compare three approaches side by side (prompt tests, model detection plus manual blocking, Bot scanning plus human review) and spell out the cost and the risk of false positives in each.
Timing and verification: spans the within-24-hour and 24–48 hour ranges. vista8's extension is untested, so detection accuracy and false positives are unknown; tualatrix's approach is personal practice. This is a follow-up: the 10-21 English brief covered Simon Willison soliciting prompts for reply bots, and the increment here is Chinese authors' home-made tools.

11. A "one prompt, four models" comparison test of MiniMax M3.1¶

Original posts and engagement:
@Fei2411｜original post｜10-20 23:50 Shanghai｜24–48 hours｜25,000 views, 188 likes
Content and what people are focusing on: the same prompt compared MiniMax M3.1 (max), DeepSeek V4.1 Flash (high), GPT 5.6 Sol (xhigh) and Fable 5.1 (medium); his ranking is Fable 5.1 > M3.1 > V4.1 Flash > 5.6 Sol. The M3.1 launch was covered in the 10-20 daily report.
Why it is worth doing / possible angles: instead of relaying a ranking, teach readers to run their own "one prompt, four models" comparison: same prompt, fixed reasoning level, explicit scoring criteria. What readers take away should be the method, not someone else's conclusion.
Timing and verification: 24–48 hours. Value observation. A single personal test on a single prompt, not a general conclusion; comparing different reasoning levels horizontally is itself biased, and the post should say so.

12. Every's "How Codex Happened": why Codex became a native desktop app¶

Original posts and engagement:
@dotey｜original post｜10-21 11:43 Shanghai｜within 24 hours｜32,000 views, 195 likes, 214 bookmarks
Content and what people are focusing on: notes from dotey's reading of the long Every article: behind the Codex App are Tibo and Andrew Ambrosino; at a time when chat built into the IDE was the fashion, Tibo decided to build a native desktop application. Bookmarks (214) exceed likes (195), suggesting readers wanted to keep it for later.
Why it is worth doing / possible angles: a product-decision retrospective — most teams at the time chose to put AI inside the IDE, and Codex did the opposite with a native app. You can look back at that decision against Codex's current shape (caching, task cards, multi-machine work) and ask whether it was the key move.
Timing and verification: Value observation, within 24 hours. Based on a relay of the Every article; the original was not read this round, so the names, timeline and assessment of the competition at the time all need verification, and it is worth reading the original before drafting.

13. Making Codex on two computers collaborate on a move¶

Original posts and engagement:
@turingou｜original post｜10-21 21:39 Shanghai｜within 24 hours｜25,000 views, 48 likes
@turingou｜original post｜10-21 11:32 Shanghai｜within 24 hours｜23,000 views, 79 likes
Content and what people are focusing on: the Mac Studio's 1TB disk filled up, so he had the Codex instances on the Mac Studio and the Windows gaming PC collaborate, moving migratable content to Windows for at least 8 hours every night. The other, more basic post: Codex automatically creates projects based on session relevance, so you do not have to pick a folder first.
Why it is worth doing / possible angles: treating the agent as an executor across machines rather than a faster editor is a use many people have not tried. You can write up your own move: how to divide the work, how to verify, and what must not be handed over.
Timing and verification: Value observation, everything within 24 hours. Personal practice, not independently verified; "at least 8 hours every night" is the author's description, with no data on actual time or success rate.

14. Making videos automatically with Claude Code: capture, animation, editing in one chain¶

Original posts and engagement:
@xiaohu｜original post｜10-21 11:29 Shanghai｜within 24 hours｜19,000 views, 172 likes
@Gorden_Sun｜original post｜10-20 18:18 Shanghai｜24–48 hours｜12,000 views, 141 likes
Content and what people are focusing on: in the video version of Xiaohu's AI daily, capture, animation, production and editing are all done by Claude Code; once the template is fixed it can be automated, and the next step is solving audio. Gorden_Sun instead went for "wish-and-it-renders", producing a Beijing 3D city-history explainer video, and included the prompt.
Why it is worth doing / possible angles: templating is the key — first break one episode into reusable steps, then let the agent run. You can write about "how to turn a one-off demo into a pipeline that runs every week", which is more useful than showing off another video.
Timing and verification: Value observation, one item within 24 hours and one at 24–48 hours. Both are personal demos without time or cost data; "next step is solving audio" is the author's next plan, not a capability that already exists.

15. A learning prompt that even Opus 5.5 got "annoyed" by¶

Original posts and engagement:
@vista8｜original post｜10-22 00:23 Shanghai｜within 24 hours｜10,000 views, 156 likes, 208 bookmarks
Content and what people are focusing on: the prompt asks for knowledge in a field that meets these conditions: no more than 50 basic units, clear combination rules, learnable in one day, and used at least once a week; then for each item, give the basic units, the combination rules and an exercise you can verify immediately. Bookmarks (208) exceed likes (156).
Why it is worth doing / possible angles: it can be turned directly into a "learn it in a day" series: pick one concrete field, run it once, and publish the output together with your own verification process, which is more convincing than relaying a prompt.
Timing and verification: Value observation, within 24 hours. The prompt text is usable, but the effect is the author's claim and was not independently verified; "even Opus 5.5 got annoyed" is a headline-style description and should not be treated as a conclusion.

16. MIT 6.S950 lecture 5: how much harness do you actually need¶

Original posts and engagement:
@shao__meng｜original post｜10-21 16:20 Shanghai｜within 24 hours｜6,524 views, 149 likes
Content and what people are focusing on: the lecture notes were updated; a 300-line teaching agent is used to discuss observability, controllability, and "how much harness is needed". With 149 likes against 6,524 views, the ratio is high.
Why it is worth doing / possible angles: this complements item 1 nicely — one side keeps adding plugins, the other asks how thin the harness should be. You can translate the lecture's three questions into judgments ordinary readers can use: when to add a wrapper and when to remove one.
Timing and verification: Value observation, within 24 hours. Only shao__meng's relay was read; the original notes and the course page were not opened this round, so the details of the 300-line teaching agent remain to be verified.

This issue has 16 items, ordered by creative value; items 11 through 16 are treated as value observations, with smaller engagement or smaller increments. Coverage gap: no video footage was watched this round, and some figures come from posts' relays and are flagged as unverified in the cards.