Awesome Hacker News: the sites that mine its comments for picks

On 31 October 2026, Andrej Karpathy put up HackerTube, a YouTube-style site that shows only videos Hacker News (HN) commenters keep recommending, and its HN thread reached 2,770 points and 613 comments¹. Within 48 hours, copies covered books, Wikipedia, products, papers, blogs and podcasts, and two directories tracked them^10,11,2,3. Below is every site as of 2 November, plus seven added on 4 November, with link, thread, method and size. The idea is 15 years old, and its problems nearly as old.

Every site runs the same funnel, from public archive to static page

HackerTube, HackerBooks, HackerZon and HackerBlogs start from open-index/hacker-news, a Parquet copy of the HN archive on Hugging Face whose card describes five-minute live updates^7,12,5,13,14. It stopped updating on 23 August 2026⁶. The live public copy is the ClickHouse Playground, a public SQL (structured query language) endpoint, which WikiHack queries^15,16. HackerProducts adds the Algolia HN application programming interface (API), and HackerPapers uses Algolia alone^17,18.

The steps vary in detail but not in order:

1. Extract candidates: links (YouTube, Wikipedia, Amazon, arXiv, blog hosts) and, for books and products, unlinked names^{7,12,16,13,14,17,18}.
2. Count distinct people: an item stays only if several commenters mentioned it^7,10,16,14.
3. Label the stance: a large language model (LLM) decides whether each comment recommends, cites, jokes about or criticises the item; WikiHack and HackerPapers describe no LLM step^7,12,16,17.
4. Fetch metadata and write a card: YouTube, Open Library, Amazon or Wikipedia supply details, and an LLM summarises what HN said^{7,12,13,16,14,18}.
5. Ship a static site with no accounts; “For You” lives in the browser’s localStorage^7,12,14.
6. Publish the data, often with a DATA.md file written for artificial intelligence (AI) agents^7,14.

The last step closes a loop. HackerTube’s about page tells readers to “give the zip to your AI agent” and ask for a variant⁷. One video creator says he did so in “1 minute 55 seconds … about 60 cents” (a single source that runs a lead magnet)¹⁹. Data engineer Simon Späti reproduced the link-counting core of HackerTube, HackerBooks and WikiHack in DuckDB²⁰.

HackerTube, HackerBooks and WikiHack set the template

HackerTube: 6,454 videos HN keeps coming back to

hackertube.org · HN thread: 2,770 points, 613 comments, 31 October¹.

Karpathy’s trigger was YouTube’s “enshittified wall of slop”⁸. The site scanned every HN comment from 2007 to 2026 for YouTube links, which “turns up well over a hundred thousand different videos”⁷. A video stays if “at least three different people linked it with some real commentary, in at least three different months”, leaving roughly 7,000⁷. Claude judged each comment’s stance, and videos rank by “how many separate months people recommended it in”⁷. The result: 6,454 videos from 39,834 labelled comments in 150 topics⁷.

A 17.8 MB data zip ships under the Open Data Commons Attribution licence (ODC-By 1.0)⁷. The credits name Claude Opus 5.5 in Claude Code, with cards “written by teams of Claude Sonnet 5.5 agents that Claude Opus coordinated”⁷. Karpathy says the build took about two hours; his only manual steps were a YouTube API key, a virtual private server (VPS) and a domain^8,21. The data stops where open-index does: the newest mention was posted on 21 August 2026²².

HackerBooks: 15,766 books, counted even when nobody linked them

hackerbooks.org · Show HN: 1,760 points, 286 comments, 1 November¹⁰.

Rohit Shinde “had Claude build one (Opus 5.5)” the day after HackerTube^10,12. Books are rarely linked, so the pipeline works on names: regexes and light natural language processing find candidate titles, Open Library resolves them, and every title is then searched across the corpus “so books are counted even when nobody linked them”¹². Claude Sonnet 5.5 then labels each mention recommend, neutral, critical or not about the book¹².

The about page lists 15,766 books, 281,363 mentions and 57,773 commenters; the Show HN had said “~16.5k books”^10,12. Bookshop.org and Amazon buy links are disclosed affiliate links¹². HackerBooks credits only open-index, yet its mentions run to 1 November 2026, from an unstated source^12,23.

WikiHack: 76,074 Wikipedia articles in a vertical feed

wikihack.netlify.app (now also wikihack.org) · Show HN: 324 points, 97 comments, 2 November¹¹.

Isaac Gemal, who made WikiTok, a feed of random Wikipedia articles, built a TikTok-style scroll through the articles HN commenters link^16,24. An article is listed when at least two different commenters linked it, or when it was itself submitted and scored 50 or more points¹⁶. Mentions inside a thread about that article don’t count¹⁶. Its about page describes no LLM step; topics come from Wikipedia’s own article-topic predictions¹⁶. It covers 76,074 articles, 540,037 links and 67,628 commenters, updated 1 November¹⁶.

The spin-offs cover products, papers, blogs and podcasts

All counts are self-reported, as of 2 November 2026 unless a section says otherwise.

hackerproducts.pages.dev · posted as a comment in the HackerTube thread¹.

HackerProducts is a prototype built from 3,768,623 comments (November 2025 to October 2026), pre-filtered by phrasing such as “I use” and “switched to” to 347,262 candidates¹⁸. Claude Sonnet has labelled 75,185, of which 34,513 name a product¹⁸. Against Claude Opus on 500 comments, Sonnet scored 87.8% F1 (a score combining precision and recall) and Haiku 78.9%¹⁸. The Opus-written verdicts reject “any clause without a citation to a real comment”¹⁸.

HackerTools: 116 developer tools

shouvik12.github.io/hackertools · posted as a comment in the HackerBooks thread¹⁰.

HackerTools lists developer tools “that Hacker News commenters actually recommend”, ranked by distinct recommenders: 116 tools and 18,114 recommendations in ten categories²⁵.

HackerZon: 9,532 Amazon products, no affiliate tag

hackerzon.pages.dev · Show HN: 2 points, 2 November¹³.

HackerZon keeps every comment that links an Amazon product page (books excluded) and counts each HN account once; links carry no affiliate tag¹³. Of 9,532 products, 8,482 were still in Amazon’s catalogue on 2 November and 1,050 were gone¹³. Its own filter shows how long the tail is: of the 8,482 products found, 7,807 were linked by a single HN user¹³.

HackerBlogs: 9.4K blogs that at least 15 people shared

hackerblogs.app · Show HN: 4 points, 2 November¹⁴.

Ian Hsiao pulled 10.7M links from 518K people out of the 49.3M-item archive, and kept 35K sites that at least 15 people shared¹⁴. Claude Sonnet then sorted out the personal, company and group blogs: 9.4K, with data to 21 August 2026¹⁴. The about page is frank that “it counts links, not praise”¹⁴. It was built with “Orbs with AmpCode” and exports follows as OPML (Outline Processor Markup Language)¹⁴.

HackerPapers: 28,636 arXiv papers

hackerpaper.kaipfister.de · Show HN: 2 points, 2 November¹⁷.

Kai Pfister’s HackerPapers pulls HN data from the Algolia HN API and ranks arXiv papers by distinct commenters: 28,636 papers, 3,323 of them linked by at least two people (the Show HN title says 3,314)¹⁷.

hnrepos: 9,335 repositories with ten or more mentions

hnrepos.com · Show HN: 11 points, 2 November²⁶.

hnrepos ranks code repositories by the number of HN discussions that mention them: 9,335 repositories on five code hosts, led by golang/go (1,294 discussions), rust-lang/rust (1,290) and torvalds/linux (795)²⁶. Its builder keeps only repositories with at least ten mentions, “since the long tail is long”, pulls the data from Google BigQuery, and splits the site into about 400 static chunks for Cloudflare’s cache²⁶. It went from start to launch in about a day, with Codex running GPT-6.1 as the agent, one of the few sites in the wave not built with Claude²⁶. Each repository is tagged by topic, project type and language, with a sparkline of its mentions over time²⁶.

Hacker Radio: 104 podcast episodes

sunmer.github.io/podcast-radio · Show HN: 2 points, 2 November²⁷.

Hacker Radio tunes between HN-recommended podcast episodes like a radio²⁷. It is a proof of concept: 104 episodes from 112 HN recommendations²⁷.

Still Unbuilt: the 100 tools HN keeps asking for

stillunbuilt.shiptiffin.app · posted as a comment in the HackerTube thread¹.

Still Unbuilt inverts the idea: it finds what HN wishes for. It collected 63,300 comments and posts with phrases like “I wish there was” or “I’d pay for”, kept the 12,441 that ask for software with an AI classifier, and ranked the top 100, for $6.00 of AI labelling²⁸. Its sharpest finding: something matching 81 of the 100 has already launched on Show HN, “1,423 times between them, and people still ask”. Its verdict: “Building isn’t the gap. Finishing is.”²⁸.

HackerChurn: the 101 SaaS tools HN complains about most

hissingkitty.com/hacker-churn · Show HN: 70 points, 24 comments, 3 November²⁹.

HackerChurn flips the stance label from praise to complaint. It reads a sample of HN comments from the last three years that name software tracked by HissingKitty, a software review site, has an AI label them, and lists a tool once it has three complaints from two threads³⁰. On 4 November it listed 101 tools from 1,551 negative comments, led by LastPass (49 complaints), CrowdStrike (47) and Evernote (44); Jira was 27th with 22³⁰. A separate tally reads “139 built their own replacement”³⁰. Data comes from the HN search API, counts update nightly, and every row now links to the comments it counted, added after a commenter asked for sources^30,29.

It was the best received of the seven sites added in this update, and its thread repeated the oldest objection in this post: “Popular services get more complaints”²⁹. Another commenter checked the first CrowdStrike complaint and found the app “confuses a negative connotation of a post, with a mention of a product inside that post”²⁹. The builder answered that the ranking is “based on % of negative vs. positive sentiment”²⁹. HissingKitty’s site blocks headless browsers, so the thread stands in for a screenshot.

hackerjobs: a year of hiring threads, parsed by Haiku

flaviaouyang.github.io/hackerjobs · Show HN: 6 points, 3 November³¹.

Flavia Ouyang turned HN’s monthly “Who is hiring?” and “Who wants to be hired?” threads into a filterable job board: role, pay, location, remote or not, technologies and industry, with a link to the original post on every card³². HackerTube and HackerBooks prompted her to give an old personal parser a front end; it is a static Astro site under the MIT licence³¹. Heuristics weren’t good enough, so Claude Haiku 5.5 parsed all 8,231 posts since January 2026 for about $12³².

Her write-up gives the rare full bill for one of these builds: $86.53 of Claude Code (Opus 5.5) over 2 hours 6 minutes of API time and almost 9 hours on the clock, for 7,486 lines added³². She “felt like a babysitter”, stopping Claude when it started coding, designing or parsing before she had approved the plan³². She also points to prior art: Jordi Noguera’s HNHIRING^31,32.

HackTok and two TV apps: new front ends on HackerTube’s data

hacktok.org · HN: 1 point, 2 November³³.

HackTok scrolls through HackerTube’s 6,454 videos one at a time³³. Two TV clients reuse the same picks: an open-source Apple TV app with a backend that refreshes the data, sideloaded via Xcode (Show HN, 5 points)³⁴, and a SmartTube fork for Android and Fire TV (comment)¹.

trustmebro.reviews/r/hackernews · Show HN: 9 points, 2 November³⁵.

davidzweig made it “a couple of months ago” and re-posted it into the wave³⁵. Gemma 4 26B, a mixture-of-experts model, did the extraction “running on pairs of NVLinked 3090 cards with vLLM, in my basement”³⁵. It holds 13M opinions on 48k products from Reddit and HN; the HN page ranks 3,223 entries from 68,607 mentions in the last 12 months³⁵.

Also posted: eight feeds, counters and front ends

HakTok (HN, 2 points) is a TikTok-style feed of HN stories, not mined from comments³⁶. Hacker GitHub (HN, 1 point) counts 384,444 links to 44,192 repositories linked at least three times³⁷. HackerNewsTV (HN, 3 points) is a daily YouTube playlist of new HN videos³⁸. HackerHack (HN), a joke that links to the HN front page, drew about 180 points³⁹.

On 3 and 4 November four more arrived, none of them mining comments for picks. HN Analytics (HN, 3 points) is a dashboard over the live ClickHouse Playground, 50.5M items, that its author calls “Fully vibe-coded slop”⁴⁰. HackerByte (HN, 2 points) summarises HN stories and GitHub Trending projects and flags AI stories with “potential”; the Show HN says it analyses comments, while the site calls its labels “an AI pre-assessment based on the title”⁴¹. HackerX (HN, 2 points) is a Chrome extension that docks the HN front page on x.com and opens live X posts for each story⁴². HackerReels (HN, 4 points) is a pre-launch app from wile.ai that turns each viral HN post into a short animated video, behind a sign-up wall⁴³.

The idea dates to 2011, and two directories now track it

In February 2011, HackerBooks.com (unrelated to today’s hackerbooks.org) mined HN and a StackOverflow dump for Amazon links and drew 351 points in its Show HN⁴. A March 2011 Wayback capture shows “6124 books quoted on Stack Overflow and Hacker News” on topic shelves; the archive is all that survives^44,2. hackeragain.com counts 22 such projects from 2011 to 2026, 11 of them book sites².

Four generations compared: data access, labelling and build effort changed; the “mention is not a recommendation” problem was only addressed in 2026, and survival is the open question.

Hackers & Store (2013): books, apps, movies and links

hnstore.co (dead; Wayback capture, April 2013) · Show HN threads: store, 190 points; movies, 114; links, 3^45,46,50.

raimonds built the 2026 recipe minus the LLM, using the “HNSearch API … to get submissions and comments” and Amazon’s API for metadata, with editions mapped to one parent id⁴⁵. Items last mentioned over a year ago were hidden “to make TOP items a little bit more varied and current”, a fix for the staleness HackerTube drew in 2026^46,1. By 2016 the domain served a WordPress apparel shop, and the links page was apparently never archived while live^48,49.

Hacker News Books (2021): deep learning for unlinked titles

hacker-recommended-books.vercel.app · Show HN: 1,359 points, 20 September 2021⁴⁷.

Hacker News Books fine-tuned BERT (Bidirectional Encoder Representations from Transformers) on about 4,000 hand-labelled comments to catch book titles nobody linked⁴⁷. Run over the HN BigQuery dataset, it found 40k comments mentioning books; the author claimed “9-50X more mentions” than earlier sites⁴⁷. It is still online, with no monetisation⁴⁷.

Before the wave: an HN video client (2025) and OrangeCrumbs (2026)

hn-yt.netlify.app · Show HN: 12 points, 13 April 2025⁵³. orangecrumbs.com · Show HN: 140 points, 14 June 2026⁵⁴.

“A Hacker News client that only shows YouTube videos” lists HN stories that link videos⁵³. Months before the wave, OrangeCrumbs already listed the Wikipedia articles HN shares, evergreen videos shared on HN, a Hall of Fame of resubmitted classics, the most-cited HN comments and the top xkcd comics⁵⁴.

hackeragain and HackerHype: directories of the aggregators

hackeragain.com (HN, 2 points) and HackerHype (HN, 6 points), both 2 November^2,3.

hackeragain lists 22 projects in six categories and marks 9 of them as unverified (7) or archived (2); its author uses the page to look for a job². HackerHype lists 26 and more, grouped by media type³.

Most of the old problems came along

Stance labels answer the oldest complaint, that a mention isn’t a recommendation, raised in 2013 and 2021^46,47. But builders measure accuracy against a larger model, not people: HackerBooks’ Sonnet labels agreed with Opus on 85% of 400 mentions, and HackerProducts’ 87.8% F1 is also against Opus^12,18. At launch, HackerBooks’ author warned that “praising a book and trashing it both count”, and a commenter noted that Mein Kampf had been “recommended” by 651 HNers¹⁰. Counting complaints has the mirror-image problem: HackerChurn’s commenters found a gloomy thread’s mood credited to a product it merely named, and rankings that favour whatever is most used²⁹.

Matching still misfires (Vanity Fair the magazine, Angry Birds)¹⁰, and HN’s taste is a fixed input. One commenter found “almost all of the top recommended books are written by men”, with about 4 of the first 100 on the home page by women¹⁰. Commenters under an Arabic-language YouTube video complained that HackerTube has no Arabic videos⁵⁵.

By design, ranking by distinct months favours old evergreen items⁷. One HackerTube user found “Everything is 6 months old or older”¹, and the data agrees: no video was first linked after May 2026²². Two things cause it: a video needs mentions in three separate months to qualify, and its data ends on 21 August, so September and October mentions never counted^7,22.

Durability is the other inheritance. A commenter who bookmarked earlier book sites later found “the only thing I got is 404”; the 2011 HackerBooks.com author replied, “Pleading guilty … I planned to revive it later, but life got in the way”¹⁰. One builder says Google came after his DEF CON video site under YouTube’s terms, and it’s now down¹. Once HN mentions drive traffic, “people will start to game your thing”; the distinct-people and distinct-months thresholds are the current defence^1,7.

The best case for the wave came from an Ask HN thread about the “flood”: “people don’t want original content from AI - they want original perspectives on existing content”⁵⁶. The best case against, from the same thread: “These apps are not a signal that you are good at anything”⁵⁶.

Our research agent, an AI model, coded 281 comments toward the whole family of sites: 270 from 10 HN threads and 11 from 2 YouTube videos, 55 of them from the 2013 and 2021 lineage threads⁹. For HackerTube and HackerBooks it read only top-level comments (150 of the 587 and 150 of the 285 served when fetched; HN’s counters show 613 and 286)⁹. Overall, 40.6% (114 of 281) were positive, 15.7% (44) negative, 25.3% (71) mixed and 18.5% (52) neutral⁹.

Later threads were more critical. On the HackerTube thread, 54.9% (45 of 82) were positive and 7.3% (6) negative. In four later threads (WikiHack, HackerHack, the Ask HN about the flood and hackeragain), 27.6% (16 of 58) were positive and 34.5% (20) negative; 44 of the 58 came from WikiHack’s thread⁹. Those threads drew different commenters, and the Ask HN is a complaint by its title, so this compares audiences, not minds changing.

The praise is about escaping the algorithm, not the technology. Algorithm fatigue came up 20 times, 15 of them positive: the site “feels like a walk in the woods”¹. Content quality came up 19 times, 16 positive: “100% pure nerd content”¹. Some dissent: “my YouTube feed isn’t trash”¹.

HN’s own taste was the most-discussed theme (28 mentions, 9 positive, 6 negative, 10 mixed): “Do you guys really trust the taste of users on here? I don’t”¹⁰. The late-wave negativity came from copycat naming (16 mentions, 8 negative) and low effort (12, 11 negative): Show HN “has now entered into its fast food slop age”¹¹. Prior art came up 18 times; of the 2025 video client, one commenter wrote, “I guess his name wasn’t Karpathy”^1,53.

The agent-built angle split commenters (17 mentions: 7 positive, 3 negative, 4 mixed); one argued that “the lowered barrier to making software is helping to untangle us from software monopolies”¹. Data export drew 9 mentions, 5 positive and none negative⁹.

What the sample can’t tell you: commenters aren’t users, and HN rewards novelty and punishes repetition. All 14 comments plugging a commenter’s own spin-off were coded positive, which inflates positivity⁹. Reddit was unavailable, and 11 YouTube comments are too few to compare platforms. The sample closed on 2 November, so the threads added in the 4 November update, HackerChurn’s included, are not in it.

What’s still unbuilt, and what to watch

As far as we found, nobody had built these requested verticals by 2 November: movies and TV (“HackerMovies someone with free tokens please”); music; games mentioned on HN; a general most-mentioned-links index like hnstore’s; Stack Exchange questions; non-English content; and a Model Context Protocol (MCP) server over the aggregators^{10,11,50,2,55}. Users also want the long tail, “books mentioned exactly twice”¹⁰.

If you build one, start from the ClickHouse Playground, which is live, or from open-index in DuckDB if data to August 2026 will do^15,6,20. Threshold on distinct people, publish the data, and publish a human-labelled sample to go with the model-versus-model accuracy figures HackerBooks and HackerProducts report^12,18.

Revisit in the first quarter of 2027. The test is whether HackerTube, HackerBooks and WikiHack still update, given one commenter’s report that “every vibe coded software I’m following has stopped updates after about a week”¹¹. Also watch whether open-index resumes; its card gives no notice⁶. hackeragain.com and HackerHype track what’s live^2,3.