A little more
about the shelf.

What’s actually in this collection?

We searched 2006–2026 Ask HN titles for “book”, “reading”, “read”, and “novel”, using monthly time windows before November 3, 2026 (UTC). We selected 395 relevant discussions across the date range and sampled up to 100 eligible comments per thread using a stable hash of each comment ID. The expanded collection retains every comment from the original pilot.

This is a sample of reading discussions, not a complete index of the date range or all Hacker News. “Less-mentioned” and all recommendation counts refer only to this sample. Thread selection favors active discussions and can miss quieter niches.

How the recommendations work

Luna receives the comment, the thread title, and parent context. It identifies books, separates recommendations from criticism and passing mentions, and extracts reading reasons and caveats. Each displayed excerpt must match text in the source comment. Luna can misclassify a reading report as an endorsement; inspect the linked comments.

Each HN username counts once per book, even if they recommend it several times. Only positive and qualified recommendations contribute to rankings. Books with uncertain identities stay out of the public shelf. A series remains a series.

Topics and niche labels are model-inferred. Reading reasons describe the commenters’ views, not an independently verified assessment of a book. Typographic jackets are visual placeholders, not original book covers.

Less-mentioned books have 2–5 distinct recommenders here. One person’s picks are shown separately. We don’t interpret a low count as evidence that a book is good or obscure.

A look at the processing

Model: gpt-6-luna, through a ChatGPT subscription. 1390 completed requests used 21,672,295 input and 3,628,348 output tokens. The run consumed subscription allowance; no dollar cost was reported. Browsing this site makes no model calls.

345 extracted references failed the exact-excerpt check; 1842 had uncertain identities. They remain in the saved records and are excluded from recommendation counts until reviewed. 5755 neutral, negative, or unclear references do not count as endorsements.

Review uncertain book identities ↗

The project includes the collection manifest, extraction responses, prompts, usage records, and a pilot report with quality findings and scaling considerations.

Our notable readers

Our editable list starts with accounts that have 10,000 or more karma in a saved profile snapshot. It can also include people we add manually. Their endorsements receive the same weight as every other reader’s.

The sampled conversations