What's this about?
A tool that pipes Claude 5's output through a separate model acting as an editor, with a prompt targeting specific tics: weird subject-and-verb combinations, subjects that should be objects, very roundabout reasoning peppered with pseudo-epiphanies, a distracting beat to the flow, and self-promotion. HN's 209 comments turned into a shared catalogue of model verbal habits.
What exactly is a "model tic"?
Every model has statistical preferences that show up as verbal fingerprints — words it reaches for, sentence rhythms it falls into, rhetorical moves it repeats. These aren't errors; each instance is fine. They only become visible in aggregate, and once you notice one you cannot stop noticing it. That's why the complaint is about texture rather than quality, and why prompting struggles to fix it: you're asking a model to avoid its own most probable outputs, which is a request it can honour for a paragraph and then drifts back from.
What's the community saying?
- The shared vocabulary of tics is the thread's real content. One commenter maintains a CLAUDE.md banning "carries," "hold," "spells," "sitting," and "where" used in place of "when" — except in SQL.
- "Claudish to English" — a rival project's name that several preferred, and a sign this is now a small category rather than one person's irritation.
- A recognizable feature request: the page needs a before-and-after example, which is what everyone actually wants from a tool like this.
- "I hope Anthropic does a post-mortem on the strange behavior their models have been displaying recently." Several reported switching to Codex over prose style rather than capability.
- A useful distinction: Opus 5 was described as far more verbose than Fable 5 — the complaint is model-specific rather than about Claude generally.