X Chinese Topic Brief¶
Translation
English translation of the Chinese original.
Today's Top Picks¶
Pick 1 | Follow-up: Fable 5.2 builds a 3D action game in one afternoon — what's reusable is that open-source skill, what got spent is 4.5 billion tokens¶
Why it's worth doing now: the Fable 5.2 game-making post was already covered two days ago (see the November 12 brief); this round adds two new things. 歸藏 (Guizang) @op7418 posted a 3D action game built in one afternoon with Fable 5.2, saying the model, scenes, action design and optimization were all done by AI, with the whole game packaged into a single HTML file of about 30MB, using 4.5 billion tokens (987 likes, 219K views, 866 bookmarks, 236 replies) — the most widely circulated Chinese post this round. The game page states "made by AI with guizang-game-skill," which means the method has been packaged into a skill anyone else can install. On the other side, Orange AI @oran_ge had just one line: "Fable 5.2 is a money pit."
Best angle to add: tell "the result" and "the bill" together. The 4.5 billion tokens is the author's own figure; at Anthropic's published Fable 5.2 list price ($8 per million input tokens, $40 output, see the November 10 daily verification), even counting everything at the input price it comes to roughly $36,000. This is only a rough calculation: in practice much of it is cache reads, and it may also run against subscription quota; the author has not published the actual spend. The more useful question for ordinary readers: how many tokens does it take to reproduce a small game with an off-the-shelf skill, and how do you draft with a cheap model first and let Fable finish.
Format suggestion: quote the original post, then add an explanation of "how to calculate the cost of demos like this" and "what a skill is."
Status: ready to draft; the game page and skill attribution have been opened and checked, the 4.5 billion tokens and "all done by AI" are the author's claims, and the cost is an estimate that must not be written as actual spend.
Pick 2 | "What a software engineer should do now": the technical moat is getting shallower; what to build up is product, marketing and "people problems"¶
Why it's worth doing now: 宝玉 (Baoyu) @dotey fully translated Scott Bolinger's short essay "What a software engineer should do now" (505 likes, 23,230 views, 513 bookmarks — bookmarks roughly equal to likes). The English original itself has 1587 likes, 152,811 views, 1706 bookmarks. The essay's argument: with every new model, pure technical skill depreciates a little more; the author increasingly feels like "a product manager managing AI," and writing user stories for the model works better than writing a rigorous technical plan; his response is to go learn product positioning, creativity, psychology and marketing — things with "no standard answer and no test set."
Best angle to add: pair it with a counter-direction reminder. Jiayuan Zhang @jiayuan_jy wrote around the same time that what most needs guarding against now is "using AI to do very fast things that never needed doing" (137 likes). Put the two together and you can write: for engineers, turning toward the judgment of "what to do" matters more than turning toward "what new skill to learn"; for non-programmer readers, the same applies to their own work.
Format suggestion: quote 宝玉's translated post and write your own trade-off checklist; or an independent original post.
Status: ready to draft. This is an opinion piece; credit the original author Scott Bolinger when quoting; the line in the text that "AI can never understand people" is the author's opinion and should not be presented as a conclusion.
Pick 3 | One person with one Agent, one person managing many Agents, a group of people with a group of Agents: the tools actually fall into three layers¶
Why it's worth doing now: answering "what's the difference between Multica and Orca, T3Code, Herdr," Jiayuan Zhang @jiayuan_jy offered a three-layer split (214 likes, 29K views, 58 replies): an individual plus a single Agent uses things like Claude Code, Codex, Pi; an individual running many Agents at once and managing a pile of sessions uses Orca, T3Code, Herdr; multi-person plus multi-Agent team collaboration uses Multica. Note: his bio says he is working on Multica, so he is an interested party.
Best angle to add: the split itself is useful to ordinary readers, helping them judge "which layer do I need now — there's no need to jump straight onto a team platform." We checked Multica's positioning against Multica's GitHub repository: it treats Agents such as Claude Code, Codex and Cursor as team members, assigns tasks and reviews results on a kanban board, with the Agent running on your own machine and self-hostable. Another post by the same author says "the harness is even less of a moat" — worth discussing together: if the shells differ little, what should you look at when choosing tools.
Format suggestion: make a three-layer comparison chart, as an independent original post.
Status: ready to draft; Multica's positioning has been verified, the classification of the other products is the author's own claim, and it's best to open each product's website before writing.
Candidate Topics¶
11 items were selected this round; no padding to 20.
1. Follow-up: Fable 5.2 builds a 3D action game in one afternoon, and the companion skill is now public¶
Original posts and traction: - 歸藏 @op7418 original post, 2026-11-12 09:43, 987 likes, 218733 views, 866 bookmarks, 236 replies. - 歸藏 @op7418 original post, 2026-11-12 10:09, 129 likes, 25310 views ("web games seem to be coming back into our lives"). - Orange AI @oran_ge original post, 2026-11-12 22:43, 10 likes, 6868 views ("Fable 5.2 is a money pit").
This item corresponds to "Today's Top Picks 1." Playable page: guizang-game-skill demo page.
Timeliness and verification: within 24 hours; follow-up (the November 12 brief already covered the Fable 5.2 game wave; this item adds the skill and the token usage). The page has been checked; the token count and cost are explained in Pick 1.
2. Scott Bolinger's "What a software engineer should do now," translated into Chinese¶
Original posts and traction: - 宝玉 @dotey original post, 2026-11-13 03:28, 505 likes, 23230 views, 513 bookmarks. - English original: Scott Bolinger @scottbolinger original post, 2026-11-12 23:12, 1587 likes, 152811 views, 1706 bookmarks. - Contrasting view: Jiayuan Zhang @jiayuan_jy original post, 2026-11-12 03:03, 137 likes, 9657 views.
This item corresponds to "Today's Top Picks 2."
Timeliness and verification: the 宝玉 and Bolinger posts are within 24 hours, the Jiayuan post 24–48 hours. The original essay was posted as an image; the correspondence between the translation and the original has been confirmed based on the citation relationship.
3. Multica versus Orca, T3Code, Herdr: a three-layer split of Agent tools¶
Original posts and traction: - Jiayuan Zhang @jiayuan_jy original post, 2026-11-12 16:42, 214 likes, 28971 views, 58 replies. - Jiayuan Zhang @jiayuan_jy original post, 2026-11-12 19:28, 28 likes, 5649 views ("the harness is even less of a moat").
This item corresponds to "Today's Top Picks 3." Fact source: multica-ai/multica repository.
Timeliness and verification: within 24 hours. The author is an interested party at Multica and the classification carries his position; say so when writing.
4. Muse video enters the Chinese creator scene: 4 open-source skills, and the "sign up on the web" discussion¶
Original posts and traction: - @cheery9998 original post, 2026-11-12 14:08, 451 likes, 17241 views. - ahhhhfs @abskoop original post, 2026-11-12 17:57, 180 likes, 53085 views, 157 replies.
Content and focus: @cheery9998 recommended 4 GitHub skills built around Meta Muse video, covering cinematic prompts, storyboarding and character consistency, an "AI virtual film crew," and batch automated generation, and noted himself that the 4th uses a third-party API, not Meta's official interface. ahhhhfs said he had finally got to use Muse and advised signing up on the web rather than in the app; that post has 157 replies (the replies were not read for this brief).
Why it's worth doing / possible angles: make clear that the name "Muse" actually covers three things: the Muse app (muse.ai, Meta's personal Agent), the Muse Video model, and Meta's developer API. According to the Meta Help Center, Muse can be logged into with a Meta account on the web or in the app, free but with a usage cap; according to Meta's July Muse Image and Muse Video launch blog, Muse Video was only a preview at the time. The Muse app is currently live only in some countries and regions and does not cover mainland China. Good to write as "what Muse video can do, where the official entry point is, and what the risks of third-party APIs are."
Timeliness and verification: within 24 hours. Do not write about sign-up methods that bypass regional restrictions; the functional descriptions of the 4 skills are the original post's claims and the repositories were not opened one by one; "Meta currently has no official Muse Video API" appears only in third-party blogs — check Meta's developer documentation before writing.
5. Follow-up: Chinese hands-on tests after the Gemini 4 Hydrogen launch — fast, and some think it's better than Carbon¶
Original posts and traction: - @huang18825 original post, 2026-11-12 21:48, 105 likes, 14685 views. - Orange AI @oran_ge original post, 2026-11-12 19:03, 70 likes, 22087 views. - Gorden Sun @Gorden_Sun original post, 2026-11-12 10:27, 17 likes, 8634 views (a same-prompt comparison of Carbon and Fable 5.2).
Content and focus: @huang18825 compared Hydrogen (straight from the API) and Carbon (run inside Antigravity) with the same prompt, felt Hydrogen did better, taking 4 minutes 47 seconds; Orange AI said Carbon is clearly stronger than DeepSeek Flash and writes more naturally, with pricing only in the Sonnet tier. Hydrogen's launch and pricing were covered in today's daily; here we only look at Chinese users' hands-on feedback.
Why it's worth doing / possible angles: "is a fast-tier model good enough" is more practical for most non-programmers. You can use the figures already verified in the daily (Hydrogen at $1.50 per million tokens input, $7.50 output) against Carbon's early-access price ($2 input, $10 output, becoming 4/20 after November 30, see the November 11 daily) to write a piece on "which tasks use the fast tier and which are worth paying for Carbon."
Timeliness and verification: within 24 hours; follow-up. Both are single personal comparisons, and the two run in different environments (API versus Antigravity), so they cannot be treated as benchmark conclusions.
6. 郭宇 (Guo Yu): having built a dozen or two vertical personal assistants, AI is like Pinduoduo — bringing in people who can't use computers¶
Original posts and traction: - 郭宇 @turingou original post, 2026-11-12 17:24, 81 likes, 18844 views.
Content and focus: 郭宇 said he has recently built a dozen or two vertical personal assistants, one of them for a Japanese hot-spring travel company, covering an all-Japan hot-spring map and ryokan sales. His observation: local people find SaaS hard to understand and use only phone and email, and AI instead lets these people who can't use computers and phones be served too.
Why it's worth doing / possible angles: most AI discussion assumes users know how to use software; this post is the opposite direction — what AI replaces is the step of "learning to use software." Good for readers running small businesses or local services: which customers have all along been shut out by SaaS.
Timeliness and verification: within 24 hours. The case is the author's own account; the assistants' concrete form and results are not public, so don't write it as verified.
7. Enterprise AI adoption: the data foundation decides RAG results; then talk about the "meta-Agent"¶
Original posts and traction: - yan5xu @yan5xu original post, 2026-11-12 09:53, 61 likes, 6541 views. - Yangyi @yangyi original post, 2026-11-12 16:20, 107 likes, 6703 views.
Content and focus: yan5xu recalled an executive project done for a large enterprise where the most basic RAG plus Agent worked well, because the financial BI data base tables were modeled to standard with complete annotations; swap in a pile of unorganized heterogeneous Excel files and "not even the King of Heaven could help." Yangyi gave three steps for enterprise AI adoption: first build trust through training, then interface the core operations of the company's various internal SaaS systems as tools for the Agent, and finally build a "meta-Agent" that automatically generates Agents per scenario.
Why it's worth doing / possible angles: the two posts tell the same story from both ends: the model is not the bottleneck, the data and tool interfaces are. Can be written as a "check these three things before adopting AI" piece for SMEs.
Timeliness and verification: within 24 hours. Both are personal experience; client names and results have not been publicly verified.
8. An open-source third-party client for the Anker recording bean, with 向阳乔木 adding local speech recognition using Fable 5.2¶
Original posts and traction: - 向阳乔木 @vista8 original post, 2026-11-11 22:49, 100 likes, 21718 views, 162 bookmarks. - 向阳乔木 @vista8 original post, 2026-11-12 09:16, 100 likes, 12768 views, 82 replies.
Content and focus: first he found that the Anker recording bean (安克录音豆) has an open-source third-party client; the next day he had Fable 5.2 redesign that client, adding a Qwen 0.6B local speech recognition model and Doubao speech transcription, with the next step being to turn it into a real-time voice input device.
Why it's worth doing / possible angles: after buying hardware, using AI to rework its software — for non-programmers this is a very concrete case of "what AI coding can do." Can be written as "giving your little piece of hardware a better app."
Timeliness and verification: the first post is 24–48 hours old, the second within 24 hours. The open-source address is in the comments of the original post and was not obtained this time; confirm the repository and license before writing; the rework's results are the author's own account.
9. "The 10 hottest GitHub repositories this month": from making slide decks and voiceovers to a Claude Code skill pack¶
Original posts and traction: - @0xluffy_eth original post, 2026-11-11 20:48, 175 likes, 12844 views.
Content and focus: the list runs from simple to hardcore: VoiceStudio for local voice cloning and voiceover, ppt-master for turning PDF or Word into editable slide decks, andrej-karpathy-skills for setting rules for models, claude-mem for recording Agent history, and so on. The post gives no stated methodology; the numbers in parentheses are presumably star counts.
Why it's worth doing / possible angles: only a few of them are directly usable by non-programmers (slide decks, voiceover, drawing), so pick those out for a "open-source tools you can use without writing code" piece. One detail already verified along the way: andrej-karpathy-skills belongs to the multica-ai organization, whose organization page shows about 232K stars, higher than the 218K written in the post — which shows the list's numbers are a snapshot from some point in time.
Timeliness and verification: 24–48 hours. The star counts and functions of the other repositories are the original post's claims and were not checked one by one; open the repositories to confirm license and maintenance status before recommending.
10. Value watch: 图拉鼎 (Tulading) has compiled 335 X accounts that "reply with AI"¶
Original posts and traction: - 图拉鼎 @tualatrix original post, 2026-11-12 10:59, 90 likes, 13651 views.
Content and focus: he said his list of "X accounts that reply with AI" has reached 335, most of which he has blocked, and he made the list public for people with the same problem.
Why it's worth doing / possible angles: AI-generated replies are diluting the quality of discussion on Chinese X, and this is something every poster runs into. You can write about "how to spot an AI reply" and "whether authors should respond"; the traction is modest but it hits close to home.
Timeliness and verification: within 24 hours. The criteria for the list are personal judgment; it does not mean all these accounts are definitely auto-replying.
11. Value watch: PiClaw — a personal assistant web workbench built on the Pi Agent core¶
Original posts and traction: - 小墨同学 @xiaomovps original post, 2026-11-11 23:39, 41 likes, 7156 views.
Content and focus: the author introduces the open-source project PiClaw: rather than rewriting the Agent core, it builds a complete web workbench on top of Pi Agent, implementing Personal Agent-like features through plugins.
Why it's worth doing / possible angles: after Pi was repeatedly mentioned in the November 11 and 12 briefs, secondary development "using Pi as the base" is starting to appear. Suited to readers interested in the build-your-own personal Agent route; traction is average.
Timeliness and verification: 24–48 hours; value watch. The project's features are the original post's claims and the repository was not opened for verification.
Interaction numbers are an API snapshot from the morning of 2026-11-13; a single snapshot cannot prove a trend.