HackerBooks: Using AI to Surface 16,500 Books Repeatedly Mentioned on Hacker News

In tech communities, book lists are never in short supply. What's scarce are "books that have actually been discussed." A developer has launched HackerBooks on Hacker News' Show HN section: an index site cataloging roughly 16,500 books, each of which must have been mentioned by at least two different HN commenters, spanning from 2007 to the present, with every book traceable back to the original comments that mentioned it.

From HackerTube to HackerBooks: A Community Idea, Reapplied

The project's origin is straightforward. The author saw Karpathy's HackerTube—which reorganizes HN discussions into a browsable format—and wanted to do the same thing for books.

One detail worth noting: the way the entire project was built is itself a signal. The author explicitly states that the site was almost entirely written by Claude, including the data pipeline and frontend, with the author mainly handling direction and decisions, plus buying the domain. The model version is listed as Opus 5.5. This "human steers, AI implements" division of labor is becoming increasingly common in individual developer projects.

Like HackerTube, HackerBooks isn't positioned as a recommendation system but as a retrieval layer for community memory. The difference is that it processes nearly twenty years of comment corpus, requiring solutions to entity matching and deduplication—difficulty on the data side rather than the interface side.

Hacker News (HN) is a tech community operated by Y Combinator, which has accumulated massive amounts of comments since launching in 2007—high information density but loosely structured. Show HN is a section within HN specifically for showcasing personal projects, where developers publish their work and receive community feedback. HackerTube is an experimental project by Andrej Karpathy that reorganizes HN discussions about videos into a browsable format. HackerBooks borrows this approach but swaps the subject to books and adds the filter of "mentioned by at least two different commenters." What these projects share: public data sources, processing costs concentrated in data cleaning, and lightweight product form.

The Data Threshold: Why "At Least Two Commenters" Matters

The most discussion-worthy design decision in HackerBooks is its inclusion criterion: a book must have been mentioned by at least two different HN commenters.

This threshold separates "someone casually dropped a book title" from "multiple people in the community independently mentioned the same book." The former is extremely noisy; only the latter has weak-signal value—it means the book was recalled in at least two different contexts. According to the author, this yielded roughly 16,500 books. The author also emphasizes that the data can be downloaded from the about page, which is important for those doing follow-up analysis.

But the author also openly acknowledges the method's fundamental flaw: it only counts mentions, without judging sentiment. Praising a book and trashing a book count exactly the same. This means the popularity ranking reflects "discussion density," not "reputation."

Traceability: Every Book Links Back to Original Comments

Every book entry in HackerBooks links back to the comments that mentioned it. This design turns a "book list" into an "argument chain."

Typical book list sites give you a ranking, and you can't tell where it came from; HackerBooks lets you click through to see the context—who mentioned this book in what discussion, whether as a recommendation, a counterexample, or an analogy. For someone trying to decide whether a book is worth reading, the original context is often more useful than the ranking.

This is also the fundamental difference between it and rating sites like Douban or Goodreads: the latter's value lies in rating aggregation, while the former's value lies in context restoration.

A Fully AI-Generated Project: Where Are the Quality Boundaries?

The author is very candid in the post: "Claude wrote almost all of it." This raises two issues—issues that这类 projects commonly face.

The first is entity matching accuracy. A book may have different editions, different translated titles, abbreviations, or series names used interchangeably. The author himself says that if a favorite book isn't included, or is matched to the wrong entity, you can report it to him. This indicates that deduplication and normalization are not perfectly completed.

The second is maintenance cost. The data pipeline was generated by AI, meaning that when fixing bugs, extending fields, or adapting to changes in HN's page structure later, the original author needs to understand the code AI wrote. This is acceptable for a personal project, but would become a real burden for long-term operation.

On the flip side, the positive significance is this: a project that would originally require three work streams—data engineering, frontend, and scraping—was compressed into two actions: "specify requirements + accept delivery." For people with clear domain knowledge but limited engineering time, this is a viable path.

"Human steers, AI implements" means the developer is responsible for requirement definition, direction decisions, and acceptance, while code writing, data pipeline construction, and other concrete implementation are done by large language models. This division of labor can significantly compress development cycles in personal projects, especially suitable for developers with domain knowledge but limited engineering time. However, for long-term maintenance, the original author needs to understand the logic of AI-generated code; otherwise, fixing bugs or adapting to data source changes becomes a burden. Additionally, the reliability of AI-generated code depends on prompt quality and acceptance standards—for parts requiring high accuracy like data pipelines, manual spot-checking and testing are still necessary.

"Community Corpus Reorganization" Projects Are Forming a Small Category

HackerBooks is not an isolated case but an extension of the same idea after HackerTube: reorganizing HN—a fast-moving, high-density but structurally messy text community—into searchable, traceable, structured views.

What these projects share: public data sources, processing difficulty concentrated in entity recognition and deduplication, lightweight product form, and core competitiveness in "the design of inclusion criteria" rather than the tech stack. HackerBooks choosing "two different commenters" as its threshold is a typical criteria design decision—it determines the character of the entire dataset.

For those who want to build their own, the author provides two directly reusable things: the downloadable data and a build path that can be copied. The only remaining question is—what kind of community corpus would you like to reorganize the same way?

Background

Entity matching and deduplication are common problems in data engineering: the same book may appear in different titles, translated names, abbreviations, series names, or even misspellings, and the system needs to determine whether these mentions point to the same entity. HackerBooks needs to normalize book title text from HN comments into unified entries, which typically involves string similarity, rule-based matching, or embedding vectors. The author acknowledges that matching is not perfect and users can report errors, indicating this step still relies on manual verification. For those building similar projects, the accuracy of entity matching often determines the credibility of the entire dataset.

One-line Summary

HackerBooks catalogs 16,500 books mentioned in HN comments, each by at least two commenters, almost entirely generated by Claude.

Paragraph Summary

HackerBooks is an index site cataloging roughly 16,500 books, with the inclusion criterion that each book was mentioned by at least two different Hacker News commenters, spanning from 2007 to the present. Every book links back to the original comments that mentioned it, turning the book list into a traceable argument chain. Inspired by Karpathy's HackerTube, the project was almost entirely generated by Claude (listed as Opus 5.5), with the author handling direction and decisions. The author acknowledges the method's flaw: it only counts mentions without distinguishing praise from criticism, so popularity reflects discussion density rather than reputation. Data can be downloaded from the about page, providing a reusable foundation for follow-up analysis. These "community corpus reorganization" projects are forming a small category, with core competitiveness in the design of inclusion criteria rather than the tech stack.