The stuff Hacker News actually recommends.
Physical products ranked by the Hacker News commenters who own them, with every opinion linked to its comment.
Visit the site · Keyboards · Buy it for life · How it works
Hacker News is full of people describing gear they've used for years, like the keyboard they've typed on for a decade or the laptop they regret. Those opinions are buried across millions of threads.
HackerBuys reads those comments and asks three questions: which product is this, did this person actually use it, and what do they think of it? It then ranks products by distinct owners, so one loud commenter counts once.
- Rankings four ways: most loved, most owned, recommended lately and long-haul.
- Verdicts adjusted for sample size: 93% from 170 owners can beat 100% from 19.
- Sources for everything: each opinion links to the original comment.
- Ownership details: years owned, whether it's still going, and which products people moved to or from.
- Tags from the data: Buy it for life, People switch to this, Divisive and Skip some generations.
- Instant search and navigation: a static site with no server, and nothing you type leaves your browser.
 Product pages: verdicts, ownership length and where people moved. |
 Categories: with sortable columns. |
 Search: matches product types, so "standing desk" also finds Fully Jarvis and Uplift. |
|
Extraction runs newest first, and the site updates as it progresses.
| As of November 6, 2026 | |
|---|---|
| Comments scanned | 828,493 of about 41 million |
| Complete years | 2022 to 2026 (through November 4), among comments that match the product filter |
| Second pass | 18,783 comments from 2023 to 2026 that the filter rejected but that name a known product and carry an opinion signal |
| Earlier years | A sample chosen to favor likely product opinions |
| Product mentions | 177,261 |
| Ranked products | 2,886 with 3 or more owners, plus 1,534 model pages |
| Models | gpt-6-luna for 2023 to November 3, 2026; Claude Haiku 5.5 for 2022 and the daily updates |
| Updates | Daily, through a GitHub Actions job |
flowchart LR
A["Hacker News archive<br/>plus recent days"] --> B["Keyword filter<br/>about 4-5% of comments"]
B --> C["LLM extraction<br/>product, verdict, ownership"]
C --> D["Results by comment id<br/>resumable by any model"]
D --> E["Build<br/>rank, tag, render"]
E --> F["Static pages<br/>on Cloudflare"]
- Collect: the
open-index/hacker-newsarchive through August 2026, then later days from the ClickHouse public demo. - Filter: regular expressions keep comments that mention a product category or brand. This keeps the cost manageable, but it misses opinions that use none of those words.
- Extract: a model reads comments in batches and returns validated JSON: brand, line, model, type, verdict, first-hand use, years owned, pros, cons and a verbatim quote.
- Rank: the Wilson lower bound over first-hand verdicts, counting each owner once, with neutral verdicts left out.
- Publish: each product and category becomes a real HTML page, with JavaScript adding sorting, search and instant navigation.
Extraction is automated and makes mistakes, so treat the site as a starting point. Every comment links to its source.
On 400 labeled comments, scored against labels from Claude Sonnet, the current settings reached about 88% precision and 82% recall, with a margin of about ±7 points. About 97% of extracted quotes appear word for word in their comments.
- Comments that pass the filter but aren't clearly reviews are the noisiest, so rankings rely on first-hand verdicts.
- A blind audit of 150 random kept mentions per model, judged by Claude Sonnet, found the first-hand verdicts that drive rankings correct about 84% of the time for Haiku (2022) and 80% for gpt-6-luna (2023), within ±8 points. Mentions by non-owners are right only about a third of the time for both, and Haiku keeps more of them.
- Rising and Fading tags stay hidden until 2019–2026 are all fully scanned, because comparing years with different coverage would mislead.
Requires uv and Python 3.12+. Raw data is about 15 GB and isn't in this repo. See scripts/01_funnel.py and scripts/02_enrich.py for the filter stage.
uv sync
# Fetch missing days, newest first (the ClickHouse demo allows 60,000 rows an hour)
.venv/bin/python scripts/21_backfill.py 2026-08-22 2026-11-03
# Queue candidates, newest year first
.venv/bin/python scripts/22_build_queue.py data/extract/queue.parquet
# Extract: stop any time, and rerunning continues where it left off
.venv/bin/python scripts/23_extract.py --adapter codex --model gpt-6-luna --workers 8 --max-credits 50
.venv/bin/python scripts/23_extract.py --adapter claude --model claude-haiku-5-5 --effort low --workers 16 --min-year 2022 --max-year 2022
.venv/bin/python scripts/26_progress.py
# Build the site (this also refreshes corpus/archive from the local queue), then preview it
.venv/bin/python scripts/14_build_site.py
npx wrangler dev
# Test
.venv/bin/python -m unittest tests.test_extract_lib tests.test_site_render tests.test_corpus
node --test tests/site_sort.test.cjsThe Codex adapter re-reads your balance before each batch. It stops below --floor, at --max-credits, if the balance rises, or on a spending limit.
The claude adapter runs Haiku through the Claude Code CLI on a logged-in subscription, with no tools, settings or MCP servers. Thinking can't be turned off there, so a batch takes about 75 seconds; use more workers. On the 400 labeled comments it reached 79% precision and 97% recall, against 88% and 80% for gpt-6-luna, so it keeps more borderline comments. Three failed batches in a row, for example at a usage limit, stop the run.
Pushes to main deploy to Cloudflare Workers automatically.
.github/workflows/daily.yml runs scripts/31_daily.py every morning. It fetches each finished day after the last one covered, queues it with scripts/30_daily_queue.py, extracts it with Haiku, rebuilds the site and pushes, which deploys. It works from a plain checkout because everything the site needs is in corpus/:
corpus/archive/<month>.parquet: text and details of the comments already processed from the full local queuecorpus/totals.json: candidates per year in the full queue, and its last daycorpus/daily/<day>.parquet: the queue for each day after that
The workflow needs a CLAUDE_CODE_OAUTH_TOKEN repository secret, made with claude setup-token. Leftover comments, for example after a usage limit, are picked up the next day.
The 15 GB archive and data/extract/queue.parquet stay local and are only needed to process older years. After a local extraction run, build the site to refresh corpus/archive, then commit it.
scripts/ filter, queue, extraction and site build
site/ the published site: generated pages plus app.js, sort.js and style.css
extract/ extraction results, one JSON file per batch
corpus/ comment text for the site build and the daily queues
analysis/ feasibility report, evaluation sets and labels
tests/ Python and JavaScript tests
Early findings are in analysis/REPORT.md.
- Comments by the Hacker News community, through
open-index/hacker-news(ODC-BY) and the ClickHouse public demo. - Inspired by HackerBooks and HackerTube.
- Not affiliated with Y Combinator.
Code is MIT. Hacker News data keeps its own license and belongs to its authors.
Made by shellbear.