archiewood/claudeisms

This project is a lighthearted, data driven look at the words Claude Code, an AI coding assistant, uses too often. The author spent several months chatting with Claude Code and had local transcripts saved automatically, so they analyzed those chat logs to see which words showed up far more often than normal. First they compared their own messages to Claude's, but that did not work well since much of what they typed was pasted log data full of technical IDs like request IDs and session IDs. Next they compared Claude's word choices to a general internet dataset, but that got skewed by ordinary programming terms like eslint or yaml. Finally they found a better comparison: a public dataset of 1.72 billion words from Stack Overflow comments written by real people between 2008 and 2020. By ranking words based on how much more often Claude used them compared to Stack Overflow commenters, the author built a list of what they call claudeisms. The results are funny: words like load-bearing showed up thousands of times more often in Claude's messages than in ordinary programmer writing. Other frequently overused words include gating, dedup, decisive, verdict, scaffolds, and handoff. The full list of 144 words, each used at least 20 times more often than typical Stack Overflow writing, is saved in a CSV file included in the project. The author admits the whole analysis was written with help from Claude Code itself, and that the results are shaped by their own personal habits and the tools they use, so it is not a scientific study of how every AI model writes. They also note that load-bearing appears directly in Claude Code's own system instructions, which explains why it shows up so often. This is a small side project built for fun and catharsis rather than a tool meant for others to install or run, though the code and data are shared for anyone curious to explore or extend the analysis themselves.