X English Hot Topics Brief¶
Translation
English translation of the Chinese original.
Top Picks Today¶
- Anthropic's Fable 5.2 prompting guide: tell it "this codebase is yours" (corresponds to candidate 1). It is worth doing now because this post by Simon Willison got about 284,000 views, 4,015 likes and 3,972 bookmarks within 24 hours, the official guide is already online, and the content can be checked point by point. Yesterday's brief covered "Fable plans, Opus executes" and tier costs; the increment here is the specific wording given officially and the measured ratios. The best supplementary angle is "when you switch to a new model, delete the old prompts first": stop writing "don't think too much" and adjust the tier instead; "never" rules need to be narrowed. Suggested format: a standalone original post with a Chinese checklist. Checked against the official documentation; ready to draft.
- A table of "AI writing ailments", 820,000 views and 21,000 likes (corresponds to candidate 2). It is worth doing now because this is one of the most-bookmarked posts in the English-speaking sphere in this window (about 15,000 bookmarks), and there is also an evaluation circulating that compares AI writing against principles from writing textbooks. The best supplementary angle is a "cross-check yourself" piece for Chinese writers: sentence patterns such as negating first and then revealing, or sticking a "this is important" label on details, are equally common in Chinese AI text. Suggested format: a standalone original post, or a quote post with Chinese example sentences. The contents of the table image were not read item by item in this round; specific entries are to be verified.
- Simon Willison gave his computer "ears" with Claude Code and found the thing that had been beeping in the storage room (corresponds to candidate 3). It is worth doing now because it is a "build a piece of software for one small thing in life" story that an ordinary person can follow: several phones and tablets listening together, locating the source to within about 40 centimeters, and finally finding a carbon monoxide alarm that needed a battery replacement. The best supplementary angle is "agents can sense the real world": microphones and cameras can both become its inputs. Suggested format: a standalone original post with links to the screenshots in his blog. Checked against the blog; ready to draft.
Candidate Topics¶
1. Anthropic releases the Fable 5.2 prompting guide: state clearly who the code belongs to, delete "don't think too much"¶
Original post and traction: @simonw's original post, November 10, 13:42, 4,015 likes, about 284,000 views, 3,972 bookmarks, 378 replies. @moritzkremb's ten-point summary, November 10, 20:14, 217 likes, about 22,000 views, 363 bookmarks. @Voxyz_ai's configuration post, November 11, 01:00, orchestrating with Fable 5.2 and having Opus 5.5 and Haiku do the work according to the official model-selection guide, 404 likes, about 25,000 views.
Content and points of interest: One point Simon singled out is this: if you write "This codebase is yours" in the prompt, Fable 5.2 is more willing to merge duplicate code and fix adjacent problems along the way, rather than only adding things on top of the existing structure. The summary post condenses the guide into ten points, including adjusting the tier by task, deleting "don't think too much", letting it finish the job, narrowing "never" rules, limiting sub-agents, and making reports readable on their own.
Why it is worth doing / possible angles: Yesterday covered division of labor and cost; today one can write "when you switch to a new model, clean out the old prompts first". The three most practical points for individual users of tools like Claude Code and Grok Bot: the tier matters more than the prompt; old cautious rules accumulate layer by layer and make it ask you about everything; stating clearly who the code belongs to changes the scope of what it touches. Also flag the cost: with an ownership sentence written, it will change more, write more tests and documentation, so you need to read the diff and not just the summary.
Timeliness and verification: Within 24 hours, a follow-up to the Fable 5.2 release. Verified: the official prompting guide confirms that with the ownership sentence added, about half of small coding tasks merge duplicate code, versus fewer than one in ten without it; output tokens are about 1.5 times as many; at medium tier and above, phrases like "just answer directly, don't think deeply" cannot stop it from thinking; the "finish the job" prompt roughly halves the number of sessions that end before completion, with output tokens increasing by about 30%. The ten-point summary is a secondhand condensation; Voxyz's configuration is a personal practice.
2. AI writing ailments: one table gets 820,000 views, one evaluation says Claude Code's replies are the hardest to read¶
Original post and traction: @MoritzSchauer's original post, November 10, 15:04, the original text is just one line, "the most complete list of LLM writing ailments I've seen", plus a table, 21,184 likes, about 823,000 views, 14,969 bookmarks, within 24 hours. @zhizdev's evaluation post, November 10, 03:25, 1,821 likes, about 118,000 views, 1,364 bookmarks, 24–48 hours old.
Content and points of interest: zhizdev's thread says AI writing often hides the protagonist and the action (the subject and predicate are not real people or things), throws together random noun strings, and has no sense of priority; all of this violates old principles in English writing books; in their evaluation, AI writing violates them far more than people do, and Claude Code's replies are the worst. The author says in a reply that prompts alone can hardly fix it, and that they are training a fine-tuned model called "English teacher". Among the discussion samples, @garybernhardt argues that the ailments you can name are only surface wording, and that the deeper patterns can be felt but not articulated (52 likes).
Why it is worth doing / possible angles: Chinese writers use AI to revise drafts every day, and this table can become a Chinese self-check list: negate first and then reveal ("not A, but B"), sticking "this is crucial" labels on details, inventing terminology. It can be read alongside candidate 1: Anthropic's own guide also specifically says "use plain words, don't invent terminology, put the key points first".
Timeliness and verification: The main post is within 24 hours, the evaluation post is 24–48 hours old. The table is an image and was not read item by item in this round; specific entries are subject to the original image; the X news page summary mentions "60% of AI text uses vague wording, versus 25% of human articles", but the original source for these numbers was not found in this round, to be verified. The evaluation method and sample size were not publicly cross-checked.
3. Simon Willison: using Claude Code to build Beep Locator, several devices listening together to find the strange noise in the storage room¶
Original post and traction: @simonw's original post, November 10, 08:09, 809 likes, about 45,000 views, 154 replies, 24–48 hours old.
Content and points of interest: According to his blog, something in the storage room was beeping intermittently, and rather than digging through boxes one by one, he had Claude Code build a locating system: a Python server coordinates, two iPads, a phone and two laptops listen simultaneously in the browser, and the location is finally pinned down to within about 40 centimeters, finding a carbon monoxide alarm that needed a battery replacement. He started with Claude Code Cloud on the phone (Opus 5.5), and switched to the just-released Fable 5.2 in the second half; Claude also found a 2007 paper on acoustic ranging on its own as the methodological basis.
Why it is worth doing / possible angles: The reader value lies in "small things in life are also worth building a one-off piece of software for", and in "giving an agent ears": it cannot listen directly, but it can write code to analyze microphone recordings, then change the code itself and keep trying. Suitable as a short piece for non-programmers.
Timeliness and verification: 24–48 hours. The process, the devices and the 40-centimeter accuracy all come from the author's blog, are a personal experiment, and were not independently reproduced.
4. Unreal Engine 5.9 preview: Epic says this is the first engine version built mainly by agents¶
Original post and traction: @TimSweeneyEpic's original post, November 11, 00:00, 2,764 likes, about 97,000 views, within 24 hours. @80Level's intro post, November 10, 16:58, 214 likes, about 16,000 views.
Content and points of interest: Tim Sweeney says the preview ports over AI acceleration features from the unfinished UE6 and adds running in the browser. The daily brief has already reported the release itself; the English-speaking sphere is more interested in the method: a small number of people directing swarms of agents working in parallel.
Why it is worth doing / possible angles: A writable angle is "how a large software company gets work done with agents": a few people directing, agents writing code, engineers reviewing and testing, assembling a preview in about a week, while acknowledging that the code still needs cleaning up. For indie game creators, running in the browser and Unreal MCP are things that can be tried right away.
Timeliness and verification: Within 24 hours, a creative angle on an event from the daily brief. Verified: the 80 Level report quotes Epic's own words, confirming that a small number of people directed agents to port UE6's AI tools and implement the Web platform, that the preview was completed in about a week, that the QA cycle was shorter than usual, that some Web renderer parts are prototypes, and that the Epic AI MegaJam is held from November 12 to 15; Epic's official release notes were not opened directly in this round.
5. patch: packaging game development tools into a CLI and MCP, with 50,000 open-source assets¶
Original post and traction: @rehan_shei's original post, November 10, 02:11, 1,696 likes, about 211,000 views, 2,718 bookmarks, 24–48 hours old.
Content and points of interest: The original post says patch puts all game development tools into one CLI and MCP, comes with more than 50,000 open-source assets, lets you mix and match other people's games, mod games on your own computer, and publish to patch. Bookmarks outnumber likes, suggesting many people saved it to try later.
Why it is worth doing / possible angles: Read alongside candidate 4: game development is turning into "letting agents call tools". For Chinese readers one can write a hands-on test: hook up its MCP with your own Claude Code or Codex and assemble a small game from the asset library.
Timeliness and verification: 24–48 hours. Verified: the patch website confirms the CLI and MCP, and 50,000 open-source assets free to use and remix, and that the desktop client is currently Windows only, with Mac requiring a wait in line. Whether the "opensource" in the original post refers to the platform itself or the assets is not stated on the website, to be verified; not tested hands-on.
6. Gemini 4 Carbon hands-on: Mollick says "a strong model that thinks differently" is useful¶
Original post and traction: @demishassabis's original post, November 11, 01:41, demonstrating Carbon reverse-engineering his early game "Theme Park", adding a DeepMind-themed project, and porting it to the web in one go, 2,015 likes, about 102,000 views. @emollick's hands-on post, November 11, 00:56, 716 likes, about 40,000 views; the same author's another post, November 11, 03:14, saying Google is back on the fast release track for frontier models, 629 likes, about 34,000 views.
Content and points of interest: Mollick says he used it in advance, that it is better than the previous generation Argon, and that having a model that thinks differently from Claude and GPT and is also strong is very useful; his demonstration was to hand a video to Carbon and have it reproduce it with three.js.
Why it is worth doing / possible angles: The daily brief has already covered the release and the benchmarks; this item only covers usage: giving the same question to two companies' models for cross-checking is more practical than only chasing "who is first". Chinese readers can watch whether it is directly usable in the Gemini app and AI Studio.
Timeliness and verification: Within 24 hours, a follow-up to an event from the daily brief. Demis's demonstration is a vendor's own claim, and Mollick's is a personal experience; a third-party post says it has already appeared in Vertex AI and AI Studio, the official availability scope is subject to Google's announcement, which was not opened and cross-checked in this round.
7. MKA: Google DeepMind and Artificial Analysis launch an agent evaluation on "reading drawings and producing deliverables"¶
Original post and traction: @ArtificialAnlys's original post, November 11, 02:22, 494 likes, about 38,000 views, within 24 hours.
Content and points of interest: The original post says MKA has 200 undisclosed real tasks covering 10 industries; the agent must read complex images such as architectural drawings, engineering schematics and medical scans, and deliver slides, spreadsheets, reports and annotated diagrams; each task is scored against an expert rubric, with about 70 criteria on average. In the first batch of results, Claude Fable 5.2 (max) leads with 57%, followed by Gemini 4 Carbon.
Why it is worth doing / possible angles: This kind of evaluation is closer to ordinary white-collar work than to doing practice questions: given a pile of files, deliver a finished product. One can write "how to judge whether AI can do this kind of work for you", reminding readers that their own files and delivery standards are the best test.
Timeliness and verification: Within 24 hours. The number of tasks, the scoring method and the 57% are all the evaluator's own statements, and the evaluation page was not opened for cross-checking in this round; the second place's score was truncated in the body text returned by the interface and was not obtained.
8. levelsio: business is getting harder to make money in, so cut costs first, move from PaaS to a VPS¶
Original post and traction: @levelsio's original post, November 10, 21:45, 2,771 likes, about 163,000 views, 3,564 bookmarks, within 24 hours.
Content and points of interest: He says moving from expensive PaaS platforms to his own VPS is usually 100 to 1,000 times cheaper, and lists the starting monthly prices of several VPS providers, recommending adding free Tailscale and Cloudflare Tunnel and closing firewall ports. This is in line with his post from the day before yesterday, "building software has never been easier, making money has never been harder" (which was covered in yesterday's brief).
Why it is worth doing / possible angles: AI makes building things cheaper, and the competition that remains is cost and distribution. One can write a Chinese explainer for indie developers on "how to launch a small AI-built product at low cost", adding the payment and regional issues that Chinese readers care about.
Timeliness and verification: Within 24 hours. The prices and "100 to 1,000 times" are the author's claims and were not cross-checked provider by provider against their websites; "closing all ports makes you safe" is a simplification, and the piece needs to add the preconditions when writing.
9. Matt Pocock's "Home Repo": one repository holds all skills and rules, letting an agent manage several projects at once¶
Original post and traction: @mattpocockuk's original post, November 11, 05:03, 1,130 likes, about 58,000 views, 1,673 bookmarks, within 24 hours.
Content and points of interest: He puts all skills, rules files and the wiki in one "home" repository, runs the agent only there, and writes in AGENTS.md where the other repositories should be cloned. The benefit is that the rules are centralized, there is only one entry point, and multiple repositories can be changed in one go.
Why it is worth doing / possible angles: Many people have an agent work on several projects at the same time, and rules files end up scattered everywhere. This is an organizing method you can follow today, and non-programmers managing multiple documents can borrow it too.
Timeliness and verification: Within 24 hours. A personal working method that needs no external verification; the effects were not tested hands-on.
10. Matt Pocock: agents amplify Goodhart's law, and teams measured on PR counts are heading for trouble¶
Original post and traction: @mattpocockuk's original post, November 10, 18:36, 464 likes, about 33,000 views, 119 replies, within 24 hours. The same author's another post, November 10, 04:33, argues that instead of just scolding an agent for a mistake, write it directly as a new lint rule, 633 likes, about 54,000 views, 24–48 hours old.
Content and points of interest: He says "when a measure becomes a target, it ceases to be a good measure", agents are especially good at gaming metrics, so organizations that measure people by PR count will run into trouble; measuring productivity was already hard, and now it may be harmful. The reply count is relatively high; the reply contents were not read in this round, so no judgment is made about the disagreement.
Why it is worth doing / possible angles: Useful for both managers and employees: when a company starts tracking AI usage, PR counts and token consumption, those are exactly the numbers that are easiest to game. It can be written together with item two as "turn one correction into a permanent rule".
Timeliness and verification: The main post is within 24 hours, the accompanying post is 24–48 hours old. An opinion piece that needs no external verification.
11. A five-word cleanup prompt: "fewer, deeper modules, smaller interfaces"¶
Original post and traction: @Voxyz_ai's original post, November 9, 21:00, 691 likes, about 49,000 views, 1,260 bookmarks, 24–48 hours old.
Content and points of interest: The author suggests that after finishing each feature, you hand Claude Code or Codex a passage: fewer, deeper modules and smaller interfaces; validate data once where it comes in and use it directly afterwards; write one-off logic close to where it is used and don't split out functions in advance; avoid try/catch when you can. The X news page also lists another cleanup prompt, "scan the whole project for broken, inconsistent and outdated things", as a hot item.
Why it is worth doing / possible angles: It can be written in contrast with candidate 1: the official guide says that only when you state clearly "the code is yours" will it proactively merge duplicate code, and this prompt is another approach. Suitable as "the routine cleanup after AI finishes writing code".
Timeliness and verification: 24–48 hours. "Line count goes down and bugs get fixed along the way" is the author's personal experience and was not tested hands-on.
12. T3 Code: any agent can now control it remotely over MCP¶
Original post and traction: @theo's original post, November 10, 08:20, 1,257 likes, about 165,000 views, 24–48 hours old. The same author's another post on November 10, 20:26 says T3 Code is popular because it is now the best coding app, 792 likes, about 68,000 views.
Content and points of interest: The original post says the flow for controlling T3 Code from other devices was redone, and pressing cmd+k and searching for "mcp" lets you set up a connector for whichever agent you want to use to control it. The X news page summary says that over the past seven days, 75% of session turns in T3 Code used Opus 5.5.
Why it is worth doing / possible angles: The brief from the day before yesterday covered how this repository lets an agent write code for itself; the increment here is "one agent directing another coding app": an assistant on your phone can remotely drive a coding session on your desktop.
Timeliness and verification: 24–48 hours. The 75% usage share comes from an X news page summary, and the original post or official data was not found in this round, to be verified; the remote control flow was not tested hands-on.
13. Value watch: Hermes agents can install other people's ready-made configurations like installing an app¶
Original post and traction: @BkashJosi's original post, November 9, 16:01, 134 likes, about 8,044 views, 213 bookmarks, 24–48 hours old. Below the usual threshold; it falls under Agent practice and is kept under the relaxed tier.
Content and points of interest: The author says he found 105 ready-made Hermes agent configurations on GitHub (personas, skills, daily tasks) that can be installed with one command, for example a research assistant, a troubleshooting assistant, and an assistant that picks a local model for this machine; he also warns that anyone can publish them, so you should read the repository before installing.
Why it is worth doing / possible angles: Personal agents are forming a "configuration store", saving the time of writing personas and skills from scratch, but security risks come with it: installing someone else's configuration means letting a stranger's rules into your agent. One can write "what to look at before installing".
Timeliness and verification: 24–48 hours. The figure of 105 configurations and the install command are the author's claims, and the repository was not cross-checked in this round; the reason for listing it as a value watch is practical value and the security reminder, not traction.
13 items in this issue.