⏱️ Lectura: 10 min
A 41-year-old historian just turned a student job shelving library books into a public semantic search tool. Benjamin Breen, a professor at the University of California, Santa Cruz, spent the summer of 2026 having Claude Code build the Book Prize Index: a database of roughly 6,500 English-language nonfiction books, filtered by a simple criterion, that won or were finalists for the language’s most important literary prizes.
📑 En este artículo
The project isn’t about generating content with AI, it’s the opposite: using AI to find quality human content amid what Breen bluntly calls AI slop.
TL;DR
- Benjamin Breen, a 41-year-old UC Santa Cruz historian, built the Book Prize Index in the summer of 2026.
- The database brings together about 6,500 nonfiction books that were finalists or winners of major English-language prizes.
- Breen used Claude and GPT-5.6 to extract and structure the data, mainly from Wikipedia.
- The tool offers semantic search: it accepts phrases like ‘classic biographies that are surprisingly weird.’
- Access is free; Breen pays for hosting and API costs out of pocket.
- The project was born as a direct counterpoint to AI slop: it uses AI to curate quality, not to generate generic text.
- Breen compares the experience to his old job shelving books in the GR 830 section of Butler Library.
Introduction
Breen’s starting point is personal. During his freshman year of college he worked as a shelver (the employee who reshelves returned books) in the A-F section of the Library of Congress classification: religion, philosophy, sociology, and history. To keep from getting bored, he adopted his own rule: open each book to a random page and read one sentence. Most told him nothing. But every so often he’d find something, like a history of the Hellenistic mystery cults or Are Clothes Modern?, and he’d end up reading several pages before putting the book back.
That method, Breen says, taught him more than any formal class, because it was filtered self-learning: the classification system and the judgment of an academic library had already selected those texts, and the fact that they’d been checked out confirmed they’d found real readers. The Book Prize Index tries to digitally reproduce that same feeling of a curated, partly random browse.
What Happened
Breen decided that the most reliable and verifiable quality criterion for English-language nonfiction books is simple: having been a finalist or winner of a major literary prize. He counted every relevant prize in the language, then asked Claude and GPT-5.6 to gather, mainly from Wikipedia, the finalist and winner lists for each one and organize them into a sortable, queryable database.
The result is a list of about 6,500 titles with enough metadata to filter and sort. On top of that base, Breen applied an embeddings model to enable semantic search: instead of looking for exact word matches, the system finds books whose meaning is close to the query.
Context and History
Breen frames the project as part of a broader decline he says he’s witnessed in his own life: that of the open, browsable-stack library, increasingly replaced by Learning Labs, group study rooms, and social spaces, while old and rare books end up in the recycling bin, replaced by digital editions.
Within that context, he argues that quality nonfiction hasn’t declined (on the contrary, he says he’s living through a golden age of the genre that’s rarely recognized as such), even as nonfiction reading falls off in the face of competition from chatbots and podcasts.
Technical Details and Performance
The central piece of the Book Prize Index isn’t the data collection (which Breen handles with Claude and GPT-5.6 reading Wikipedia lists), but the semantic search layer on top of those 6,500 titles. An embeddings model converts each book and each query into a numeric vector that represents its meaning; the closeness between vectors (not word matching) determines which results appear first.
That’s why queries like modern France or social history work, but so do much more specific phrases like classic biographies that are surprisingly weird, which the model can associate with titles no keyword search would find, like a biography of the James family (the Whartons) that Breen describes as weirder than its cover suggests.
The following diagram summarizes the full flow, from the prize lists to the final semantic query.
flowchart TD
A["Prize lists on Wikipedia"] --> B["Claude and GPT-5.6 extract the data"]
B --> C["Structured database of 6,500 titles"]
C --> D["Embeddings model generates vectors"]
D --> E["Semantic search by phrase"]
E --> F["List of recommended books"]
| Method | When to use it | Advantage | Limitation |
|---|---|---|---|
| Keyword search | Looking for an exact title or author | Precise and predictable | Doesn’t find topically related books if they don’t share words |
| Semantic search (embeddings) | Exploring by concept, tone, or similarity to another book | Finds non-obvious connections between titles | Results are sometimes puzzling or hard to justify |
| Human curation (librarian or shelver) | Serendipitous discovery within a physical space | Quality filter already validated by experts and real readers | Doesn’t scale beyond a physical collection and library hours |
How to Get Started (Try It)
The Book Prize Index is a simple case of applying embeddings to a structured collection of texts. Reproducing the idea on a small scale doesn’t require complex infrastructure: a single embeddings model and a handful of titles are enough.
from sentence_transformers import SentenceTransformer
modelo = SentenceTransformer("all-MiniLM-L6-v2")
consulta = "classic biographies that are surprisingly weird"
vector_consulta = modelo.encode(consulta)
print(vector_consulta.shape)
This first block loads a general-purpose embeddings model and converts the query into a vector. The result is a numeric array (typically 384 dimensions for this model) that represents the meaning of the phrase, not the exact words it contains.
import numpy as np
def similitud_coseno(a, b):
return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))
libros = [
"The World of Yesterday, by Stefan Zweig",
"Life & Times of Michael K, by J. M. Coetzee",
"Cosmos, by Carl Sagan",
]
vectores_libros = modelo.encode(libros)
puntajes = [similitud_coseno(vector_consulta, v) for v in vectores_libros]
ranking = sorted(zip(libros, puntajes), key=lambda x: x[1], reverse=True)
print(ranking[0])
The second block calculates the cosine similarity between the query and each book in the corpus, and sorts the results from highest to lowest semantic closeness. The book that appears first in ranking is the one closest to the meaning of the search, not necessarily the one that shares the most words with it. To check that the full Book Prize Index corpus responds well to a query, Breen himself recommends searching for a book you already know and checking whether the related suggestions make thematic sense.
💡 Tip: if you’re building your own semantic search engine over a small collection, start with a lightweight model (like MiniLM) before moving to bigger ones: compute cost grows fast and doesn’t always improve perceived relevance.
Impact and Analysis
What sets the Book Prize Index apart from most tools labeled as AI is that it doesn’t generate text: it curates text that already exists and was written by people. Breen is explicit about this: he says there’s nothing AI-generated in the project except the tool that collected and structured the data, and the semantic search engine.
That distinction matters because it separates two very different uses of the same language models: using them to produce new content (with the risk of adding more AI slop to the internet) and using them to organize and make navigable human content that already passed a quality filter, in this case, that of the judges of the major English-language literary prizes.
💭 Key point: semantic search doesn’t replace the librarian or the shelver, it replicates the curated randomness of a well-organized shelf, but at the scale of thousands of titles and without depending on physically standing in front of the right shelf.
The case of Stefan Zweig illustrates well the kind of discovery Breen is trying to enable: his memoir The World of Yesterday, about pre-World War II Vienna, shows up among the results of related semantic searches, a book Breen describes as profoundly moving, written just before Zweig took his own life in Brazil in 1942.
What’s Next
Breen mentions that, once the data was gathered, generating visualizations from it became an almost natural side experiment: the same structured corpus that feeds the semantic search also works for exploring aggregate patterns across prizes, decades, and categories. The project remains free and funded by Breen out of pocket, with no announced plans for direct monetization beyond a supporter subscription option for his newsletter.
An honest limitation: because it depends on Wikipedia’s coverage of prize lists, the corpus inherits that source’s biases, with better coverage of recent English-language prizes than of older or less documented ones. And as with any embeddings-based search, some results are, in the author’s own words, a bit puzzling, though that unpredictability is also part of what reproduces the feeling of browsing a physical shelf.
📖 Summary on Telegram: View summary
Try it yourself: install sentence-transformers with pip install sentence-transformers and run the second code block with your own list of favorite books to see how well the model captures their thematic resemblance.
Frequently Asked Questions
What is the Book Prize Index?
It’s a free, searchable database of about 6,500 English-language nonfiction books that were finalists or winners of the language’s major literary prizes, created by historian Benjamin Breen.
How is semantic search different from keyword search?
Keyword search finds literal text matches. Semantic search converts queries and books into numeric vectors and ranks results by closeness of meaning, so it can find related books even if they don’t share the same words.
Is the Book Prize Index free to use?
Yes. Breen covers the hosting and embeddings API costs out of his own pocket, with the idea that more people will discover and read good nonfiction books.
What AI models were used to build it?
Breen used Claude and GPT-5.6 to gather and structure the prize list data, mainly from Wikipedia, and a separate embeddings model to enable semantic search.
What does it mean that the project was vibe-coded?
It means Breen didn’t write the code by hand, but instead directed Claude Code to generate and iterate on it, describing the result he wanted in natural language instead of programming each function directly.
Why does Breen present it as the opposite of AI slop?
Because the project doesn’t use AI to generate new content, but to organize and make discoverable quality human content already filtered by literary prize juries, at a time when much of the new content on the internet is mass-produced by AI.
References
- Quality non-fiction books are the antithesis of AI slop: Benjamin Breen’s original article describing the Book Prize Index project.
- Sentence embedding, Wikipedia: a general explanation of how models convert text into vectors to measure semantic similarity.
- Claude Code, Anthropic: the official page for the tool used to build the project.
- Library of Congress Classification, Wikipedia: the library classification system Breen mentions as his first experience with filtered curation.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Peter Herrmann en Unsplash
0 Comments