The post is a one-screen interactive analysis of GitHub pull requests that clusters vocabulary and highlights a group of terms that have become strongly associated with Claude output. The author is careful about the claim. This is not a classifier for “Claude text” so much as an unlabeled vocabulary cluster whose frequency jumps in recent data. People broadly bought the signal anyway because it matches daily experience. Many said the site captured why Claude now feels harder to read even when it is useful, and they praised the presentation almost as much as the analysis itself.
The conversation landed on a sharper distinction than the post itself makes. Most people were not objecting to any single word like “seam,” “
sidecar,” or “load-bearing.” They were objecting to saturation and misuse. Some of these are legitimate terms from software engineering, operations, or
agile. “Seam” predates LLMs by decades and has a specific meaning in legacy-code refactoring. “
Spike” and “sidecar” also have established uses. The complaint is that Claude reaches for this register constantly, often where plain language or concrete references to files, functions, and failure modes would communicate better. That makes the prose feel like faux precision. Dense, technical-sounding, but often less informative than a shorter direct sentence.
That led to a broader diagnosis of what changed. The dominant view was not that Claude is “too intelligent” for humans. It was that reinforcement and synthetic training are pushing models toward a narrow house style that scores well somewhere in the pipeline and then spills into everything. People pointed to
RLHF,
reasoning traces, agent-to-agent summaries, and synthetic or distilled training corpora as plausible sources. Several noted that the visible shift seems to line up with newer Claude releases around spring 2026, where benchmarked capability may have improved while readability got worse. A recurring framing was that the model now writes partly for other agents or for internal evaluators, then hands users a compressed summary full of unexplained metaphors and overloaded verbs.
A smaller but important thread pushed back on treating the vocabulary itself as evidence of degradation. The better argument there was that terms like “seam” or “load-bearing assumption” can be genuinely compact and useful among humans who share context. The problem is not jargon existing. The problem is that a model has no excuse to hide behind shorthand when it can cheaply be specific. Readers also noted that this language is already leaking into workplaces and personal writing, which makes the post less like a joke list of Claude-isms and more like an early snapshot of machine-influenced office dialect.
louisabraham.github.io