Turn Hacker News into a Skill
Sorts out the top 5 posts on the front page, then saves the method as a Skill you can reuse with one sentence.
- 1Create the Skill
- 2Reuse in one sentence
帮我看下 Show HN 榜单最近热帖
最近思考
The user wants to see recent hot posts on Show HN. Let me open the page. I should check tabs first, then open https://news.ycombinator.com/show. Let me be efficient: list_tabs and open_tab. I'll open the Show HN page.
我先打开 Show HN 页面看看。
历史思考 1
1/3
There's a skill hn-top5_brief available with a tool hn_top5_brief__hn_top_posts. But that tool is for the HN front page. The user wants Show HN. Let me load the skill to see if it supports Show HN. Actually let me first read the page to see Show HN posts. But the skill tool might work on show page too since it parses top posts. Let me load the skill first.
历史思考 2
2/3
The skill is for front page. User wants Show HN. I can use the same parsing approach on the Show page. Let me try the skill tool on the Show page — it parses tr.athing, which Show page also has. Let's try it with limit 10.
最近思考
3/3
The Show HN page default ordering is not by score — it's by time? Actually Show HN page shows newest-ish ordered. Rank 1 has 94 points 13 hours ago, rank 2 has 2 points 1 hour ago. It's mixed — the /show page is sorted by... actually it's sorted by "ranking" similar to front page but for Show HN. Hmm, rank 2 with 2 points and 1 hour ago suggests it's ordered by recency more than score? Actually HN /show is sorted by ranking score which includes time decay. "最近热帖" — recent hot posts. The /show page is the default Show HN listing. But there's also "Show HN" newest: https://news.ycombinator.com/shownew (newest). The user asks "Show HN 榜单最近热帖" — recent hot posts on Show HN leaderboard. The default /show page is fine. But the ordering seems odd. Let me consider getting a more "score-sorted" view. Hmm. Given ambiguity, I'll present the /show listing (10 items) plus note that the Show HN list is HN's ranking order (time decay + votes). Should I add one-line takeaways? The skill says step 4 requires reading originals. That's 10 tabs — heavy. The user just asked to "看下" (take a look). I think a table with scores/comments and a brief overall observation is enough. But providing takeaways adds value. Given cost, maybe fetch descriptions for the top few? Let me consider effort vs value. The user request is casual: "帮我看下 Show HN 榜单最近热帖". A clean table + short summary of main themes (title-based) is appropriate. I could add brief notes based on titles, but I must not fabricate facts. I'll keep to observable data (title, score, comments, site, age) and give a light thematic read clearly derived from titles. Maybe I should try to get a bit more: fetch a couple of original pages for descriptions? That could improve quality. Let's do a moderate approach: open the top 3 originals and grab meta descriptions via run_script. Actually that's several tool calls. Hmm, but value is decent. Let me instead do one script that opens tabs? No, run_script can't open tabs (that's a navigation action). Actually tools.open_tab is available via tools proxy? The protocol says open_tab creates working tab; in run_script, "提交、导航、关闭 tab 等页面生命周期动作不要塞进 run_script". So no. I'll do a lightweight approach: give the table, and note takeaway inferred from titles. Actually let me at least fetch descriptions for the top 5 by opening tabs one by one — that's ~5 open_tab + 5 script calls. Too many. Alternative: HN item pages include the post text for Show HN. For Show HN posts with a URL, no text. I'll skip deep-reading and clearly mark that the one-line notes are title-based, not from reading originals. Hmm, but protocol prefers verified claims. Title-derived summary is fine if marked. Actually, let me do a compromise: fetch meta descriptions for the top 4 via one script that fetches? CORS blocks fetch. No. Decision: deliver table + thematic summary based on titles, explicitly labeled. Also mention there are 30 items on the page, showed top 10, and ask if they want the full list or details. Also should I save to data_store? Not necessary unless user wants. Skip. Note timing: collectedAt 2026/10/23 12:14:30 Beijing time. Let me write response with footnotes to HN discussion links? Footnote rules: claims about facts from web pages need citations. Score/data comes from HN page — I can cite the HN show page and item pages. Let me add a few footnotes for the top items with links I have (HN discussion URLs and original URLs). These are "verified" in the sense the page was read. I'll present a table with links, and footnotes for a few notable ones. Keep footnotes reasonable — maybe cite the Show HN list page and 3-4 key items. Let me build the table for top 10.
新开页面
Show | Hacker News
1/3
加载技能
2/3
hn-top5-brief
采集 Hacker News 首页前 N 条热帖的标题、得分、评论数、链接与讨论页，并提炼一句话看点
资源
plugin.json
json · 1.6 KB · 声明解析 Hacker News 首页前 N 条帖子的工具与表格渲染。
scripts/hn-top-posts.js
script · 3.0 KB · 解析 HN 首页 tr.athing 帖子行，输出结构化热帖列表。
读取 HN 热帖
3/3
HN 首页热帖
| # | 排名 | 标题 | 得分 | 评论数 | 站点 | 发布时间 | 原文链接 | HN 讨论 |
|---|---|---|---|---|---|---|---|---|
| 1 | 1 | Show HN: Uberblok – an interactive museum of ZFS internals | 94 | 23 | storagebits.workers.dev | 13 hours ago | https://uberblok.storagebits.workers.dev/ | https://news.ycombinator.com/item?id=50180257 |
| 2 | 2 | Show HN: A wysiwyg Markdown editor for macOS that works with agents | 2 | 3 | kirahowe.com | 1 hour ago | https://kirahowe.com/2026/oct/22/editmd-0-1-0 | https://news.ycombinator.com/item?id=50203836 |
| 3 | 3 | Show HN: Turn regular JavaScript into a neural net | 38 | 11 | chenglou.me | 9 hours ago | https://chenglou.me/handwired/ | https://news.ycombinator.com/item?id=50172984 |
| 4 | 4 | Show HN: UI micro interactions prompts for AI | 74 | 42 | typeui.sh | 9 hours ago | https://www.typeui.sh/micro-interactions | https://news.ycombinator.com/item?id=50199290 |
| 5 | 5 | Show HN: agent-usage-guard – a Claude Code mod to protect your usage window | 22 | 9 | github.com/rbartoli | 10 hours ago | https://github.com/rbartoli/agent-usage-guard | https://news.ycombinator.com/item?id=50198936 |
| 6 | 6 | Show HN: CATT 0.2.1: Test agents/services against a Muse/Instinct-alike, in CI | 7 | 1 | github.com/coverage-cat | 5 hours ago | https://github.com/Coverage-Cat/consumer-agent-testing-tool | https://news.ycombinator.com/item?id=50201860 |
| 7 | 7 | Show HN: Replacing CAPTCHAs with adaptive proof-of-work in WordPress | 10 | 7 | wordpress.org | 7 hours ago | https://wordpress.org/plugins/gdpr-compliant-recaptcha-for-all-forms/ | https://news.ycombinator.com/item?id=50194762 |
| 8 | 8 | Show HN: Dovepass – can I legally fly a drone along this exact path in SF? | 30 | 8 | dovepass.com | 14 hours ago | https://dovepass.com | https://news.ycombinator.com/item?id=50156779 |
| 9 | 9 | Show HN: Build a world on your phone, explore it in XR on Vision Pro and Quest | 13 | 3 | camilleworld.com | 9 hours ago | https://3d.camilleworld.com/play/ | https://news.ycombinator.com/item?id=50199349 |
| 10 | 10 | Show HN: Omnara – Open-source alternative to Claude Managed Agents | 110 | 34 | github.com/omnara-ai | 1 day ago | https://github.com/omnara-ai/omnara | https://news.ycombinator.com/item?id=50183892 |
这是 Hacker News 的 Show HN 榜单（也就是 /show 列表页）当前的排序。我按页面顺序取了前 10 条；页面共有 30 条。分数和评论数是 2026-10-23 12:14（北京时间）的快照，会随时间变化。
Show HN 榜单前 10
| # | 标题 | 得分 | 评论 | 站点 | 发布 |
|---|---|---|---|---|---|
| 1 | Show HN: Uberblok – an interactive museum of ZFS internals | 94 | 23 | storagebits.workers.dev | 13 小时前 |
| 2 | Show HN: A wysiwyg Markdown editor for macOS that works with agents | 2 | 3 | kirahowe.com | 1 小时前 |
| 3 | Show HN: Turn regular JavaScript into a neural net | 38 | 11 | chenglou.me | 9 小时前 |
| 4 | Show HN: UI micro interactions prompts for AI | 74 | 42 | typeui.sh | 9 小时前 |
| 5 | Show HN: agent-usage-guard – a Claude Code mod to protect your usage window | 22 | 9 | github.com/rbartoli | 10 小时前 |
| 6 | Show HN: CATT 0.2.1: Test agents/services against a Muse/Instinct-alike, in CI | 7 | 1 | github.com/coverage-cat | 5 小时前 |
| 7 | Show HN: Replacing CAPTCHAs with adaptive proof-of-work in WordPress | 10 | 7 | wordpress.org | 7 小时前 |
| 8 | Show HN: Dovepass – can I legally fly a drone along this exact path in SF? | 30 | 8 | dovepass.com | 14 小时前 |
| 9 | Show HN: Build a world on your phone, explore it in XR on Vision Pro and Quest | 13 | 3 | camilleworld.com | 9 小时前 |
| 10 | Show HN: Omnara – Open-source alternative to Claude Managed Agents | 110 | 34 | github.com/omnara-ai | 1 天前 |
几点观察
- 榜首与最高分不是同一条：页面上分数最高的是第 10 名的 Omnara（110 分、34 评论），但它在列表里排第 10，因为 HN 的排序综合了时间衰减，这条已经发布 1 天。列表顺序并不等于分数顺序，第 2 名只有 2 分也是同样原因（1 小时前刚发）。
- 今晚的主线明显是 AI Agent 工具：Omnara（开源版 Claude Managed Agents）、agent-usage-guard（给 Claude Code 加用量窗口保护）、CATT（在 CI 里测试 agent 服务）、editmd（面向 agent 的 Markdown 编辑器）占了 4 条，讨论度也集中在 Omnara 和 typeui 上。
- 讨论最热的不是最高分：typeui.sh 的「给 AI 用的 UI 微交互提示词」74 分、42 条评论，是这 10 条里评论最多的，说明它更有争议或更值得讨论。
- 技术向的项目更稳：Uberblok 把 ZFS 内部结构做成交互式博物馆，94 分；handwired 用原生 JavaScript 手写神经网络，38 分——这两条是典型的「实现类 Show HN」，得分稳、评论理性。
以上看点都只依据标题、分数和评论数得出，还没有逐条打开原文。需要我：
- 逐条读原文（README、正文、meta 描述），给每条补一句有依据的一句话看点，并保存到 data_store 表；
- 或者把 Show HN 全部 30 条内容拉成完整表格（包含「最新」列表 shownew）；
- 或者只针对 Omnara、typeui 这两条讨论最热的帖子，抓取评论区观点摘要。
你想选哪一种？告诉我，我就继续。