X Chinese Hot Topics Brief¶
Translation
English translation of the Chinese original.
Top Picks Today¶
1. Uncle Bob's ub-mode: turning "Clean Code" into an Agent Skill¶
Why now: This item has three verifiable anchors at once — the repo really exists, the stars are climbing, and the discussion has already moved from "retweeting the skill file" to "why copying it still doesn't stop bad code." The main post went out 10-27 19:20 with 153K views, 1,367 likes and 2,006 bookmarks; bookmarks outnumber likes, the classic "saving it for later" pattern. The same author posted a long hands-on write-up on 10-28 11:37 with 17K views and 216 bookmarks, which shows the topic is continuing rather than being a one-off spike. On GitHub, unclebob/ub-mode was created on 10-25; the GitHub API showed 3,743 stars on the morning of 10-29, and the repo holds only two files, README.md and SKILL.md — that ratio of "lots of rules, very little code" is itself a writable angle.
Added angle: The core tension is that "rules written into the prompt" can't stop bad code; you need executable gates to go with them. The three tools named in the post map onto three kinds of gates: Crapper for complexity, coverage and CRAP score, Dryer for structurally similar duplicate code, Mutator for mutation testing, plus a UML Viewer. It is worth spelling out what mutation testing means to non-programmers — not "the tests passed" but "if you deliberately break the code, can the tests catch it." The second contrast is item 4's personal rule list and Karpathy's CLAUDE.md: rule files are popular because "acceptance" has become the bottleneck — many people write rules, few can automate acceptance.
Format: One real end-to-end run — take a small repo through the whole loop of "tidy the data structures first, write failing tests, then change the code, and finally run the three tools," and paste the scores the tools output; that is more useful than paraphrasing the steps. Add one caveat: the post itself warns that copying the skill file won't automatically block bad code, and the tools have to be set up separately.
Status: Ready to draft, partly unverified. The "152-line" skill file and the tool thresholds (one post says complexity must stay below 10) are claims from the post, not checked line by line; star counts change, so cite the live GitHub page.
2. AI helps someone squeeze iOS 16 onto a Xiaomi phone¶
Why now: @op7418's post on 10-28 16:17 got 132K views, 278 likes and 94 replies, pushing something that had lived in a small geek corner onto the front of the Chinese timeline. Its reach comes from one line of contrast: "the AI cracking and porting wave is absurd." What separates this round from past ROM-flashing waves is that the AI credit can be found in the first-party repos — the drivers in ios-sweet were written by Codex with Claude executing instructions, 74 of ios-topaz's last 100 commits note collaboration with Claude, and commits in HackiOS even carry Co-Authored-By Claude Opus 5.5. This is a rare checkable case of "AI compressing low-level work that used to take a team months into a few people and a few days."
Added angle: The real wall isn't "will it boot," it's closed-source drivers and firmware. ios-sweet reached the home screen on 10-18, running unmodified iPhone X iOS 16.7.16 userland, but with no GPU, no Wi-Fi/Bluetooth/cellular/audio/camera, most security mechanisms off, and it wipes Android user data; ios-topaz added Wi-Fi and GPU (Mesa freedreno), while Bluetooth and cellular are still missing. Laying these out as a "boots / usable" checklist carries more information than shouting "Android on iOS." The second layer is risk and licensing: Apple's license only permits iOS on Apple devices, and these ports switch security mechanisms off by default. The third layer is method: AI can fill in the code writing and the doc lookup, but not the baseband and charging parts that have no source.
Format: No tutorial, no flashing steps. Make it a comparison of three devices at three stages of progress, with one closing line: AI shortens the coding time of a port, not the reverse-engineering cycle of firmware.
Status: Ready to draft, with one key line to cut. "The Xiaomi 15 can already run iOS 16" is a claim inside @op7418's post and is unverified — all that could be found was UniXNU's Xiaomi 15 port repo, with no verifiable working build; the public results are the Redmi Note 10 Pro, Redmi Note 12 4G and POCO X5 Pro.
3. Claude's 5x/20x only counts the 5-hour window: weekly limits and the Duet reset card¶
Why now: This is a correction-type topic, correcting a marketing line that keeps getting retold. @Charles77xixi's post on 10-28 17:47, 35K views and 109 bookmarks, made it clear that "20x / 5x refers to the 5-hour window, not the weekly limit"; @fkysly restated it on 10-28 23:31, 21K views. The official help center wording matches: Max 5x/20x means 5x and 20x Pro's "per-session usage allowance," a session resets every 5 hours, and there is a separate cross-model weekly limit whose exact value is not published. What heavy users actually hit is the latter, which is exactly the one the official docs don't quantify.
Added angle: Write it as "pick your tier by your own usage" rather than "which tier is better value." First, look at the two progress bars under Settings > Usage — whichever bottoms out first, the 5-hour one or the weekly one, is your answer, and use ccusage to track it if needed. Second, mark the post's conversions clearly as guesses: "two Max 5x > one Max 20x," "six Pro ≈ one Max 20x," roughly $1,500–3,500 to max out 5x at API prices and $2,800–5,500 for 20x — all of these come with preconditions. Third, independent estimates contradict each other: LLM CFO puts the 20x weekly limit at about 1.5–2x that of 5x, while DevelopersIO's test puts it around 3.5x, so "two 5x is the better deal" depends on which estimate you trust and can't be treated as settled. Fourth, multiple accounts may also conflict with the terms of service.
Worth writing up alongside it: Duet — once approved in the US region you get one full reset, covering both the 5-hour and the weekly limit, valid until November 30; the entry point is Reset for free under Settings > Usage on web or desktop, and there is no button on mobile or in Claude Code. @Rakel076's 14K-view post on 10-28 13:10 and @dotey's post on 10-28 09:29 both mention it — an information gap where "those who know have already used it, those who don't are still waiting."
Format: Lead with the conclusion (what you are buying is the 5-hour window; what blocks you is the weekly limit), then give a three-step self-check, and finish with a judgement on "when 20x is actually worth it." You can attach @tualatrix's 21K-view hands-on post from 10-27 19:55 as a counterexample: on Claude Pro, one session that invoked a subagent took the 5-hour allowance from 100% to 0.
Status: Ready to draft. No verifiable source was found for "therefore sued," so don't write it; the ~85% token gross margin is an analyst's derivation from the S-1, and @dotey's post gives the method — H1 revenue of 16.286 billion and cost of revenue of 2.380 billion yield about 85.4% gross margin, with Opus 5.5 costing about $0.6/$3.0 — this must be labeled as a derivation, not a disclosure.
Candidate Topics¶
1. Uncle Bob's ub-mode: turning "Clean Code" into an Agent Skill¶
Post and traction: @longhaiqwe123 10-27 19:20, 153K views, 1,367 likes, 150 reposts, 23 replies, 2,006 bookmarks; the same author's hands-on long post 10-28 11:37, 17K views, 169 likes, 216 bookmarks; @LawrenceW_Zen 10-28 08:00, 6,914 views, 54 likes.
What it says: The post gives a 152-line skill file whose flow is "tidy the data structures first, write failing tests, then change the code," followed by checks on complexity, duplication and test effectiveness; it ships with three tools — Crapper (complexity/coverage/CRAP score), Dryer (structurally similar code) and Mutator (mutation testing) — plus a UML Viewer. LawrenceW_Zen compresses the same thing into three steps: get it running (a failing test plus a minimal change), then get it right (fix the root cause, cut complexity, remove duplication, mutation testing), and working habits (record each step and the reason, commit in small steps). The post also warns: copying the skill file won't automatically block bad code, and the tools need separate setup.
Why it is worth doing: Of this wave of "rule files," it is one of the few entries with a toolchain, an original repo and a named author. You can write "why rules in a prompt can't stop bad code," or write about which parts of old-school engineering discipline actually land once repackaged into an agent-readable format, and which parts are just a new shell.
Timing and verification: The main post is in the 24–48 hour band, the hands-on post within 24 hours. The repo is verified to exist, created 10-25, 3,743 stars via the GitHub API on the morning of 10-29, holding only README.md and SKILL.md; "152 lines" and the complexity threshold are claims from the post, not checked line by line.
2. AI helps someone squeeze iOS 16 onto a Xiaomi phone¶
Post and traction: @op7418 10-28 16:17, 132K views, 278 likes, 94 replies; the post says "the AI cracking and porting wave is absurd," that the author came across a Xiaomi 15 running iOS 16, and that they are getting ready to try it themselves.
What it says: The first-party material is more worth reading than the post. ios-sweet ports Apple's open-source XNU (the iOS 16 generation) to the Redmi Note 10 Pro, running unmodified iPhone X iOS 16.7.16 userland; it reached the home screen on 10-18, with no GPU, no Wi-Fi/Bluetooth/cellular/audio/camera, most security mechanisms off, and it wipes Android user data; ios-topaz ports to the Redmi Note 12 4G, with Wi-Fi and GPU (Mesa freedreno) already working and Bluetooth/cellular still missing; Tuanzi Tech Daily reports that HackiOS runs iOS 16.5 on a POCO X5 Pro, open-sourced on 10-22, with commit sign-offs including Co-Authored-By Claude Opus 5.5; a roundup by HaveFun says ios-sweet's drivers were written by Codex with Claude executing instructions, and that 74 of ios-topaz's last 100 commits note collaboration with Claude.
Why it is worth doing: This is material that can explain both "AI speeds things up" and "the limits of AI speedups" at once. The coding cycle really is compressed, and closed-source drivers and firmware really do block progress; add licensing and security risk (Apple's license only permits iOS on Apple devices, and the ports disable security mechanisms by default), and you can write a technical observation that doesn't urge anyone to try it.
Timing and verification: Within 24 hours. Repos and media are verified, and the AI credit is documented. "The Xiaomi 15 can already run iOS 16" is unverified: only UniXNU's Xiaomi 15 port repo was found, with no verifiable working build; the public results are the Redmi Note 10 Pro, Note 12 4G and POCO X5 Pro.
3. Claude's 5x/20x only counts the 5-hour window: weekly limits and the Duet reset card¶
Post and traction: @Charles77xixi 10-28 17:47, 35K views, 159 likes, 37 replies, 109 bookmarks; @fkysly 10-28 23:31, 21K views, 44 likes, restating the same conclusion; @tualatrix 10-27 19:55, 21K views, 55 likes, 41 replies; @Rakel076 10-28 13:10, 14K views; @dotey 10-28 09:29, 21K views, 65 replies.
What it says: The main post says 20x/5x refers only to the 5-hour window, with the official term being per-session; the weekly limit is about 3.5x and 6x that of Pro; from this it derives "two Max 5x > one Max 20x" and "six Pro ≈ one Max 20x"; at API prices, maxing out 5x costs about $1,500–3,500 and 20x about $2,800–5,500; it also mentions "therefore sued," and, citing an analyst's inference from the S-1, a token gross margin of about 85%, used to explain why banning Max accounts is so strict. Two other posts cover Duet: approval in the US region comes with one full reset. On the cost side, @dotey 10-28 11:37, 39K views, 190 likes, gives Opus 5.5 pricing at about $0.6/$3.0 and a gross margin of about 85.4% derived from H1 revenue of 16.286 billion and cost of revenue of 2.380 billion.
Why it is worth doing: The marketing language of subscription tiers is misaligned with users' actual bottleneck — a topic with long-term reuse value. You can write the "pick a tier by usage" methodology, or write the quota economics of the two companies — one with a 5-hour window plus an unpublished weekly limit, the other with repeated resets across 28 days of daily releases (item 11).
Timing and verification: The main post is within 24 hours, the tualatrix post in the 24–48 hour band. The help center Max explainer matches the post's "per-session," but the weekly limit's exact value isn't published; independent estimates contradict each other (LLM CFO at 1.5–2x, DevelopersIO at about 3.5x), so the conversion conclusion can't be treated as settled. No verifiable source was found for "sued," so don't write it.
4. Tw93's AI Coding rule list¶
Post and traction: @HiTw93 10-28 16:52, 49K views, 1,106 likes, 134 reposts, 91 replies, 1,422 bookmarks; @dotey 10-28 13:28, 23K views, 202 likes, 249 bookmarks, adding one worktree rule.
What it says: The list has seven sections — requirements, rules, coding, troubleshooting, review, release, retrospective — with 20 items in total. Key points include: shift from writing implementation details to wish-style programming, and dare to raise the bar; state the context and goals and leave fewer details to it; use the best model; ask a human first about new features and flags; cut down to a minimum usable version; teach the agent about verbal instructions ("take a look" means read-only, while committing, releasing and deleting need confirmation); put only rules that can't be seen in the code and rejected approaches into AGENTS.md; pasting keys to an agent counts as a leak; rules that must be followed should be enforced by scripts, not prompts; before writing code, say clearly what changes and what it affects; have the agent add verification scripts; have multiple agents each use their own worktree; if a bug isn't fixed after two tries, add logging, and after three, clear the context and start over; separate the agent that writes code from the agent that reviews it, with at most 3 review rounds; one-way doors like deletion, payment and release must be looked at by a human; write how to verify and how to roll back in the commit; give it to early adopters first; keep a test case for every production bug and rerun them when switching models. @dotey's addition is a rule requiring AGENTS.md to state that no worktree is left behind — delete the worktree and branch when the task is done.
Why it is worth doing: The list itself is highly readable, but the real increment is the reordering — grouping by "where a human should spend time" fits readers better than the seven process sections. It can also be set against item 1: the division of labour between personal verbal rules and executable gates, one handling judgement, the other handling interception. For non-programmers, you can pick just the three that are usable first.
Timing and verification: Within 24 hours. A personal experience list, needing no external verification; there are 20 items in total, and item 18 in the original post has a duplicated number, so say so when citing the count.
5. Anthropic's platform engineering lead: the four hard problems a thousand agents still need programmers to solve¶
Post and traction: @dotey 10-28 07:49, 55K views, 398 likes, 87 replies, 542 bookmarks, a long-form paraphrase; @snwiki238337 10-27 17:02, 20K views, 97 likes, 130 bookmarks, reading it through the lens of an engineer's identity crisis.
What it says: The subject is Katelyn Lesse, who posted on October 25 (US time). The four problems are: scheduling (agents are stateful, wait a long time, and burst; Kubernetes doesn't fit; it mentions Google's open-source Agent Substrate, which can sleep and wake in half a second); identity and authorization (telling apart whether a human is acting or an AI is acting for them, short-lived credentials, authorization is largely a blank); security (one agent tricked by a malicious instruction on a web page can pass it on to its peers or write it into shared memory, so strong sandbox isolation is needed); and monitoring with a master switch (records that agents can't tamper with, and one switch to stop the whole system and revoke permissions).
Why it is worth doing: All four problems translate into versions an individual can feel — whose account your agent uses, whose money it can spend, whether being tricked by a web page can corrupt other agents, and whether you have a one-button stop. That last question hooks straight into item 10 (giving an agent its own phone number) and item 3 (quota and authorization), so they can be strung into one set.
Timing and verification: Both posts are in the 24–48 hour band. The Intelligent Artifact summary matches the four problems and notes that these are design suggestions and open questions, not evidence that large agent teams have already produced results; Lesse's original post went up on X at 10-26 02:19 UTC with about 630K views (third-party tracking). Business Insider reported on 10-22 that she said her team's code is all written by Claude and the team is about 200 people. When writing, don't turn "design suggestions" into "practices already running in production."
6. The "one-person company" acquired by Anthropic: how many people is Mithril really (Follow-up)¶
Post and traction: @xiaohu 10-28 13:07, 31K views, 207 likes, 43 replies, 127 bookmarks; @oran_ge 10-28 07:24, 19K views, 134 likes, 56 replies.
What it says: @xiaohu says "OPC really can work," the claim being one founder plus Claude, with contracts, logistics and coordination across 11 external labs all handed to Claude, pushing a therapy to positive mouse results in under 18 months, with human trials at the earliest by the end of 2028; @oran_ge thinks it "may be of major historical significance, when everyone says OPC doesn't hold up." For first-party information see R&D World: one full-time employee, Cody Tranbarger, backed by 8 paid advisers and 11 external labs doing antibody production, binding assays, cell experiments and mouse studies; the acquisition happened in July and was disclosed on 10-27; the focus is LAMA2-related congenital muscular dystrophy, with a candidate drug planned for 2027 and human trials at the earliest by the end of 2028; Tranbarger calls the model Pipeline-in-a-Person.
Why it is worth doing: The real shape of a "one-person company" is one full-time decision-maker plus a ring of outsourced experts and contract labs, with Claude doing the coordination. You can spell out which layer it removes (administrative coordination) and which it doesn't (experiments, expert judgement, funding). That breakdown is more useful than taking sides on "does OPC hold up."
Timing and verification: @xiaohu's post is within 24 hours, @oran_ge's in the 24–48 hour band; the 10-28 X AI daily already mentioned this acquisition, so this item is a follow-up from a creation angle. The Next Web adds: the deal terms were not disclosed, the technical report has not been peer-reviewed, and mouse grip strength was about twice that of the untreated group. Don't write it as "one person plus AI made a drug."
7. GitHub flooded by "agent workbenches": IDEs, eyes and rules¶
Post and traction: @GitTrend0x 10-27 08:30, 66K views, 764 likes, 121 reposts, 1,298 bookmarks. The post is about 47 hours 50 minutes old, already at the tail end of the 24–48 hour band, so if you are going to write it, do it soon.
What it says: The post lists 5 projects: JetBrains/thinkrail (a lightweight IDE built around the pi coding agent, which the post says went from 2,854 stars to a gain of 1,156 in a day), Panniantong/Agent-Reach (lets agents read and search Twitter, Reddit, YouTube, Bilibili, Xiaohongshu and more, pitched as zero API cost), multica-ai/andrej-karpathy-skills (a single CLAUDE.md with four principles), alchaincyf/huashu-design (Design Skill 3.0), nightly-labs/openbot (a local desktop workspace giving Codex/Claude/Grok each a station and a task queue, developer preview, very broad permissions). Live star counts from the GitHub API on the morning of 10-29: thinkrail 5,024 (created in July), Agent-Reach 105,620, andrej-karpathy-skills 225,982, huashu-design 27,962, openbot 3,241.
Why it is worth doing: The summary line "agents lack a workstation, eyes and rules" expands directly into three layers: workstation (thinkrail, openbot), eyes (channels like Agent-Reach that read platform data) and rules (CLAUDE.md and the various Skills). Each layer carries its own risks, especially the latter two.
Timing and verification: Tail end of the 24–48 hour band, with traction already fading. Star counts are live API values from the morning of 10-29; the single-day gains quoted in the post are unverified. Tools like Agent-Reach that get around login walls to scrape platform data should be discussed in terms of platform terms and account risk; openbot has broad permissions, so only run tasks you trust. The 10-26 Chinese topic brief already covered Huashu-Design's birdwatching prototype, so don't repeat it.
8. "AI editing can replace junior editors now"? (Follow-up)¶
Post and traction: @dotey 10-28 11:20, 40K views, 405 likes, 133 replies, 402 bookmarks; @yaohui12138 10-28 10:26, 75K views, 426 likes, 480 bookmarks; @LufzzLiz 10-28 21:34, 13K views, 146 likes, 179 bookmarks; @Ryrenz 10-27 19:09, 6,528 views, 119 likes, 119 bookmarks.
What it says: @dotey's judgement is that after Opus 5.5, AI agents editing video can replace junior editors, cutting video and generating animation with HTML and JavaScript, analogous to when Claude Code first launched. @yaohui12138 claims "one command recreates a Douyin science video with 365K likes," and says 99% of the people around them can't use Claude. @LufzzLiz gives a workflow of a digital-human talking head plus Opus 5.5 motion effects. @Ryrenz's post is a value pick: capcut-cli lets an agent create drafts, add subtitles and apply templates for you in CapCut, updated from 0.28 to 0.31 in October.
Why it is worth doing: Take apart what a "junior editor" does — rough cut, subtitles, templates, motion effects, rhythm judgement. The first four already have tools taking over; rhythm judgement stays with humans. That lands better than arguing about "will they be replaced." The other thread is divergence: people who can't use Claude can go through the CapCut CLI or other agent-based alternative chains.
Timing and verification: The main post is within 24 hours, @Ryrenz's post in the 24–48 hour band. "Replacing junior editors" is an opinion; "one command recreates" and the 365K likes are claims from the post, unverified; the capcut-cli version numbers are also claims from the post. The 10-28 Chinese topic brief already covered Renoise MCP, so this item only covers this judgement and the split in toolchains, to avoid repetition.
9. Magpie makes "what the redaction changed" visible (Follow-up)¶
Post and traction: @yetone 10-28 14:09, 72K views, 677 likes, 150 replies, 229 bookmarks; the same author @yetone 10-28 19:38, 46K views, 292 likes, 94 replies.
What it says: The first post's origin is a friend who didn't dare turn on redaction, afraid that altering the content sent to the large model would make the agent dumber; the author says it won't degrade quality and added a display (video) in Magpie. The second post covers how Magpie plays after becoming fully plugin-based (video).
Why it is worth doing: The biggest enemy of a privacy feature isn't performance, it's that users don't dare turn it on. This one can be just one line of judgement — trust needs to be visible. Two layers underneath: the trade-off of redaction (can the model still reason after entities are replaced), and how to self-test (the same task with it on and off, comparing how much the output differs).
Timing and verification: Within 24 hours. "Won't degrade" is the author's own account, with no independent test; the video content can't be read through the interface, so defer to the original post for the actual UI. The 10-26 and 10-27 Chinese topic briefs already covered Magpie's local redaction and the gateway outage; this item is a follow-up that only writes "trust needs to be visible" and doesn't restate product features.
10. Get an agent a phone number of its own (Value pick)¶
Post and traction: @turingou 10-28 01:15, 48K views, 268 likes, 269 bookmarks; the same author @turingou 10-27 13:57, 32K views.
What it says: The first post is about giving an agent a Japanese phone number, used to register accounts and collect points; the plan is Povo 2.0, recommended by Grokbot, with a zero base fee, an eSIM issued as soon as you enter a credit card, and the My Number Card scan done in under a minute. The second is about having a product engineer on Grokbot pick the product name and register the domain — it doesn't even pop a confirmation card, "as long as your Cloudflare has a card attached, it will just charge you."
Why it is worth doing: Once an agent has its own number, email and card, it has an "identity." This is the personal-scale version of item 5's identity and authorization problem: isolation (a dedicated number, a dedicated card) is a good habit, but limits must be set when charges go through without confirmation. It could also be written as a "four things to settle before opening accounts for an agent" piece, for example a spending cap, a whitelist of sites it may register on, and a recovery path after something goes wrong.
Timing and verification: 24–48 hours; a value pick, not high traction but a rare angle. Povo 2.0's zero base fee and its signup flow are claims from the post, unverified; charges going through without a confirmation prompt is personal experience, unverified. This kind of content involves real-name verification and payments, so write only the mechanism and the risks, not the steps.
11. OpenAI's 28 days of daily releases, day 23: another quota reset¶
Post and traction: @lxfater 10-28 20:06, 47K views, 83 likes, 37 replies; @dotey 10-28 10:46, 30K views, 41 replies; @don_tech1 10-28 14:36, 8,248 views.
What it says: @lxfater's own words are "posting resets every day like a clown" and "the model is slow and dumber again," which is an emotional opinion. @dotey gives operational information: Codex resets at midnight Pacific time. @don_tech1 records that Day 23 started with an iCloud plugin stunt, while the actual action was another quota reset. For background see MacRumors, which reported on 10-27 that ChatGPT launched iCloud mail and calendar plugins; the Day 23 reset is recorded by the reset tracker limitdigest, applying to paid Codex and ChatGPT Work users.
Why it is worth doing: Quota resets have turned into a marketing rhythm, while what users read is "it wasn't enough in the first place." The best use of this item is to set it beside item 3 and write up the two companies' different ways of playing "quota economics": one turns resets into a daily show, the other hides the weekly limit and publishes only multipliers. The difference one and the same user feels on the two sides is content in itself.
Timing and verification: Within 24 hours. The reset record and the eligible users can be checked on a third-party tracker; "dumber" in the post is a subjective feeling, so don't write it as a conclusion. ChatGPT Pro/Business's up-to-$500-per-month API credits have already gone into the 10-29 X AI daily, so this item doesn't repeat it.
12. A frontend and Chinese typography Skill for removing the "AI flavor" (Value pick)¶
Post and traction: @vista8 10-27 21:07, 27K views, 204 likes, 382 bookmarks; @thinkszyg 10-27 09:54, 5,432 views, 119 likes, 122 bookmarks.
What it says: @vista8 says a netizen recommended Hallmark Skill, "so ugly it has no AI flavor," which can make Claude produce something with the Codex feel; before that they thought emil-design-eng and impeccable were the best. @thinkszyg open-sourced "Measure," aimed at the problem of applying English typography habits to Chinese — tightening letter spacing on big headlines, italic emphasis, 10px small text.
Why it is worth doing: The "flavor" of AI-generated pages comes from default design conventions, so changing the conventions is more effective than changing the prompt; Chinese typography has its own set of rules, and transplanting the English one directly causes problems. You could do an episode comparing "the same page with three different Skills," letting the visuals do the talking, which is more intuitive than explaining principles.
Timing and verification: Both posts are in the 24–48 hour band. Both are introductions from the posts, with the repos unverified; Hallmark Skill's exact origin and license were not checked. A value pick, medium traction, best placed in a softer slot.
Coverage gap: In the account sweep, three groups — lxfater/yetone/Gorden_Sun, shao__meng/AlchainHust, and DashHuang/geekbb/tualatrix — were swamped by high-frequency results without paging through; targeted searches were run as a supplement, but low-frequency posts from these accounts may still be missed.