Show HN: The load-bearing vocabulary of Claude
Points and comments are a snapshot, not live.
Scraped 461,121 GitHub pull requests to find Claude's most overused vocabulary.
The site analyzes 1,000 GitHub pull requests daily, totaling 461,121 PRs and 51,079,244 words over 595 days. It identifies vocabulary clusters, showing "load-bearing" is 39.47× more frequent in one cluster. A scrollable word cloud visualizes terms like "refusal," "premise," "byte-identical," and "vacuously." The presentation is concise and interactive, with word sizes varying for emphasis.
What commenters are saying
Commenters overwhelmingly critique Claude's obnoxious, jargon-heavy style. Many say terms like "vacuous" and "provenance" are common, though some users report different experiences. A key criticism is that Claude's explanations are inscrutable, requiring a PhD to parse, and often invent jargon on the fly ("fold a capstone into a name"). Two camps emerge: those blaming RLHF/watermarking vs. those blaming hidden chain-of-thought summaries. One commenter notes humans may unconsciously adopt LLM language quirks, like listing examples with "etc." The site author responds to design praise.