The load-bearing vocabulary of Claude: ten clusters in 461,000 pull requests

An independent data project groups GitHub pull-request descriptions by the words they are written with, rather than by any label a researcher chose in advance. The result is ten clusters, and one of them behaves unlike the rest: it made up 0.7% of the corpus at the start of 2025 and roughly 39% of it by the middle of 2026.

What the project measures

The site, The load-bearing vocabulary of Claude, takes 1,000 pull-request descriptions a day through GitHub's search API and splits them into ten groups with KL-divergence k-means, a method that compares probability distributions over words. Neither stemming nor a stopword list is applied: a word is any run of letters, digits, slashes, hyphens and underscores containing at least one letter, so tokens such as load-bearing or mutation-tested survive whole.

The published accounting covers 603 collected days, 595 of them falling in 85 complete weeks from 6 January 2025 to 17 August 2026 — 461,121 descriptions and 51,079,244 word appearances, of which 19,798 words cleared the frequency floor.

The cluster that arrived

The arriving cluster's most distinctive words are load-bearing, plainly, quietly, nobody, outright, re-derived, mutation-checked, mutation-tested, refused, vacuously, premise and restated. The project notes that this vocabulary should look familiar to anyone who uses coding agents. Across the whole corpus the cluster holds 8.2%, but over the last four weeks it accounts for about 37% of the descriptions.

There is also a detector: paste a pull-request description, or a GitHub link, and the same fit reports whether the text belongs to that arriving cluster.

How the sample is built

The methodology write-up explains why the obvious source could not be used. GitHub's public event archive stopped carrying text: since the commit array was removed in October 2025 its feed holds almost only push events. The search API works instead because its date filter accepts timestamps, so a window can be minutes wide. Ten five-minute windows are drawn each day, one per 2.4-hour block, and each returns full body text.

One limit is stated openly. A search query returns at most 1,000 results and a page holds a hundred, so the corpus samples rather than enumerates. Pull requests authored by four apps — dependabot, renovate and similar bots — are excluded inside the query itself, since they are roughly 90% of app-authored bodies.

What it does and does not show

The project does not claim to attribute any single description to a particular tool. It shows that the language of pull requests has split statistically, and that one half of that split is growing fast on a yearly scale — the kind of shift that matters to anyone who reads code review, docs and project history for a living.