Declauding v0.13.0 ships taxonomy of 52 LLM writing tics

A new release of the declauding Claude skill, v0.13.0, catalogs 52 LLM writing tics and bundles a toolchain to scrub them from drafts.

The skill, maintained in the oaustegard/claude-skills repo on GitHub, "removes LLM prose tics from a draft and returns plain human technical prose." Patterns are grouped by mechanism rather than phrase, on the premise that blocklists miss paraphrases. The release notes describe the largest cluster — entries 1 to 23, plus 37 and 38 — as a single move: "the sentence is built to make the reader feel a finding arrive, instead of stating the finding." The prescribed fix is blunt: "put a real subject in the subject slot, say the thing, stop."

Entries 43 to 47 are where the skill turns on itself, flagging what the notes call "the flat, verdict-shaped register a clean pass lands in, which is now the fastest-growing cluster of GitHub pull request descriptions there is." Outputs come in two shapes: a clean rewrite, or a self-contained HTML diff showing every edit with its original and its reason. The release also ships a linter, a diff checker that reports lost or invented content, and an embedding-based sentence ranker. Two practitioners in the directory we maintain posted the link this week.