HackerBooks creator roansh says Claude Opus 5.5 wrote almost all of the project’s website and data pipeline, turning years of Hacker News comments into a browsable book catalog.
In the project’s November 1, 2026, Show HN announcement, roansh described roughly 16,500 books mentioned by at least two different commenters, covering discussions from 2007 onward. Each book links back to the comments that mentioned it, and the site offers a downloadable dataset.
“Claude wrote almost all of it, including the data pipeline and the site,” the creator wrote, describing their own role as steering, making decisions, and buying the domain.
The project also comes with a warning: praise and criticism both produce mentions. That distinction became a product-design problem as readers examined its rankings.
The current documentation describes an additional Claude-powered classification pass to separate recommendations from neutral discussion, criticism, and references to something else entirely. The result is a useful case study in agent-built AI products: generating the application and deciding what its numbers mean are separate jobs.
From HackerTube to a Downloadable Bookshelf
Roansh explicitly credited Andrej Karpathy’s HackerTube as the inspiration for HackerBooks. The idea was to apply a similar browsing experience to books discussed on Hacker News. Zeniteq previously covered that predecessor in Karpathy Says Claude Built HackerTube in Two Hours.
The HackerBooks announcement supplies no comparable build duration. Its concrete claim concerns the division of labor: Claude generated almost all of the pipeline and site, while the creator directed the work. It provides no measured percentage or code-authorship audit.
The About page credits Claude Code and documents considerably more than a generated front end. HackerBooks extracts possible book references from an archive, resolves them against an external catalog, counts mentions, classifies their meaning, and exposes the results through browsing pages and a CSV download.
That combination makes it a more useful example than an interface demo. Readers can follow a ranking back to individual comments, while other builders can download the book-level data.
The available HN thread snapshot showed 1,841 points and 304 comments. That establishes substantial attention within this particular discussion, without making its participants representative of readers generally.
How the Extraction Pipeline Works
HackerBooks starts with a public archive of Hacker News items and comments. Its documented pipeline then uses five stages:
- Find candidate references. Regular expressions and lightweight natural-language processing identify patterns such as quoted titles, phrases containing “by Author,” ISBNs, and links to Amazon, Goodreads, or Open Library.
- Resolve candidates against Open Library. Matching supplies a stable work identifier, authors, first-publication year, covers, and subjects.
- Search the whole comment collection. Once a title has been resolved, the pipeline looks for it throughout the archive, including comments without book links.
- Disambiguate difficult titles. A small language model checks whether ambiguous names such as Dune, Influence, and Range actually refer to books.
- Rank the resulting records. Scores give more weight to distinct commenters than raw mention totals. The Rising view compares recent mentions with a book’s history.
The corpus-wide search is an important expansion step. A system that only counted explicit bookstore links would miss ordinary conversations where someone simply names a book.
It also creates an obvious precision problem. A title that is an everyday word can match unrelated sentences. A title shared with a film or game can match a different work. Resolving a candidate to an Open Library entry gives it a consistent identity, but subsequent matches still need to refer to that identity.
According to Puneet Nanda’s November 2 analysis on DEV Community, roansh described an offline Python-and-DuckDB build feeding a plain-JavaScript site with static JSON on Cloudflare. Nanda also reports that covers reside on R2 and that roughly 30 kilobytes of top-book data load before the remaining catalog.
That is a sensible architecture for an archive-derived catalog: do the expensive extraction work ahead of time, then serve prepared results. Its practical tradeoff is freshness. New discussions only appear after another data build.
The Launch Count and Current Catalog Differ
The launch post’s roughly 16,500 books should be treated as a launch estimate, not a permanent catalog size.
The About page, alongside a dataset dated November 2, 2026, lists:
| Measure | Published count |
|---|---|
| Books | 15,766 |
| Mentions | 281,363 |
| Comments containing mentions | 225,861 |
| Distinct commenters | 57,773 |
The homepage also displays 15,766 books.
These are different snapshots. The evidence does not establish which filtering, matching, or catalog changes account for the difference from the initial estimate.
The distinction between units matters, too. One comment can mention several books, and one person can mention a book repeatedly. None of those totals directly measures how many people recommend a title.
For anyone reusing the CSV, I’d keep its generation date alongside it. Comparing exports without tracking their dates could turn pipeline changes into an apparent shift in reading habits.
When a Mention Becomes a Misleading Recommendation
The creator’s launch warning was direct: favorable and unfavorable discussion both counted. The problem becomes consequential when an interface presents those counts as recommendations.
Nanda’s analysis reports that a commenter flagged Mein Kampf appearing as recommended by 651 people. According to that account, the site’s label was subsequently changed to mentioned by.
That incident is reported by the secondary source. The project’s current documentation directly acknowledges the underlying problem: mention counts can make a frequently discussed or criticized book appear to be a favorite.
The difference is more than wording. A mention count measures attention. A recommendation count makes a claim about someone’s judgment. Criticism can increase the first while providing no support for the second.
Nanda also reports that Vanity Fair the novel was picking up references to the magazine, prompting a change to require a connection with Thackeray. This illustrates a different failure: the system can identify the wrong entity before it ever evaluates whether the comment is favorable.
My view is that fixing the label was the essential first step. A sophisticated classifier can improve the dataset later; the interface first needs to describe honestly what its existing count represents.
A Second Claude Pass Adds Meaning, With Limits
The current About page says Claude Sonnet 5.5 labeled each mention using four categories:
| Label | Intended distinction |
|---|---|
| Recommend | The commenter recommends the book |
| Neutral | The book is discussed without a recommendation |
| Critical | The comment criticizes the book |
| Not about the book | The reference concerns another work or subject |
References classified as unrelated are dropped. The Most recommended page ranks books by distinct recommenders, while the mentioned-by count includes everyone who mentioned a book. A person qualifies as a recommender if they recommended that book at least once.
The site says simpler approaches, including keyword rules, sentiment scoring, and a TF-IDF classifier, were tried first and performed less accurately. It does not publish comparative measurements for those alternatives.
Frequently Asked Questions
3 questions
1Did Claude build all of HackerBooks?
Creator roansh says Claude Opus 5.5 wrote almost all of the site and pipeline. The creator steered decisions and bought the domain. No measured code-authorship percentage was supplied.
