Method · HackerFriction

The short version

A small classifier ranked every public Hacker News comment up to 2026-11-04 by how much it reads like a usability complaint. Claude Haiku read the top of that ranking, decided which comments really are usability complaints, named the product each is about and summed up the problem in a line. Claude grouped the names into products and a small model grouped similar problems into issues. Issues are ranked by how many different people raised them in each period. Justusability reviews the issues that reach the home page by hand.

Step by step

Data. The Hacker News archive mirrored by open-index/hacker-news (ODC-By), itself from the ClickHouse public dataset, topped up to 2026-11-04. Deleted and flagged comments are left out.
Finding complaints. A small classifier, trained on Claude's labels, ranked all 42.1 million comments. Claude Haiku 5.5 read the 418,532 at the top (1% of all comments), plus every comment using general complaint language such as “user hostile” or “dark pattern”, and labelled 253,188 of them usability complaints about a product people would search for. Bugs, outages, prices and policies don't count.
Products. Claude Haiku grouped the names complaints were filed under into products: iPhone takes in iOS and the iPhone keyboard; Mac takes in macOS and MacBooks; iPad, Gmail and Microsoft Teams stay separate. A name that spans two products, such as “iOS and macOS”, goes to the more specific one, or keeps a page of its own. Some fixes were made by hand. Broad categories such as “websites” or “software” get their own pages but aren't ranked on the home page. Small and local businesses are left out.
Issues. A small language model grouped each product's one-line summaries by meaning. Claude Haiku dropped groups that were only vague dislike or bug reports, reworded the rest as plain issue names, and merged near-duplicates.
Listing. An issue is listed once at least 3 people have raised it, in at least 3 different months and 3 different threads. To rank in a period it also needs 3 people and 3 threads within that period.
Ranking. By how many different people raised the issue in the period: this year (the 12 months to 2026-11-04), the last 5 years, or since 2007. So one launch thread can't carry an issue to the top on its own, each person counts in the thread of their first complaint in the period, and no thread counts for more than a third of an issue's people (or 3, whichever is more). The page shows the real number of people, so now and then an issue with more people sits just below one with fewer, spread over more threads.
Review. Justusability reviews every issue on the three home page lists: approves it, edits its name, merges it with another, or drops it if it isn't a usability problem. A dropped issue leaves the ranking and the next one moves up. No issue is removed for any other reason. Anyone can report a mistake through the “Something wrong here?” link on each issue, which opens a public request on GitHub.

Reading the counts

“N people” counts distinct Hacker News usernames whose comment was labelled a usability complaint about the issue.
The counts are a floor. On a random sample of 5,000 comments, about 3.6% of all Hacker News comments are usability complaints; the part Haiku has read holds roughly one in five of them. Counts will grow as it reads further down the ranking.
Quotes, on reviewed issues only, are short extracts of at most 25 words, copied exactly, and link to the full comment. Lines under unreviewed issues are Claude's one-line summaries, not quotes.

Limits

Hacker News isn't everyone. It skews towards developers and technical people, and towards the US. Treat these as problems people report, not measurements of how many users hit them.
Labels and groups are Claude's. Justusability spot-checked labels and reviews every issue on the home page, but we haven't yet measured label accuracy against a full hand check. Expect some noise, especially on unreviewed pages.
The data stops at 2026-11-04. Comments after that aren't counted yet, so the latest year is partial.
One person can use several accounts, and a username isn't a verified identity.

Credits

The comments belong to the Hacker News commenters who wrote them. Data from open-index/hacker-news and the ClickHouse public dataset. The idea comes from Andrej Karpathy's HackerTube, and we learned from Hacker Books and WikiHack. Built by Justusability with Claude.

Privacy