- Author
- GitHub user oaustegard, in the claude-skills repository
- Published
- 9 November 2026
- Kind
- GitHub release of a Claude skill, with a README and skill instructions
- Original
- github.com/oaustegard/claude-skills/releases/tag/declauding-v0.13.0
- Subject
- Writing with models
What it is
declauding is a Claude skill that rewrites drafts such as PR descriptions, READMEs and essays. It returns clean text or an annotated HTML diff. Version 0.13.0 adds a step that finds the draft's point before sentences are edited. The release page links a zip to install it in Claude.
What it says
- The README says it lists 52 entries, mostly grouped by mechanism and not by phrase.
- The skill says nearly every tic is one move. The sentence stages a finding instead of stating it.
- Version 0.13.0 adds step 1a. Write the draft's one point first, and rebuild if it is missing.
- The skill warns that over-editing is worse than the tics. It leaves flat prose with no opinions.
- The README describes a linter that only finds candidates. A clean report means nothing alone.
- The skill bars the rewrite from adding any fact, name, number, date, quote or citation.
- Step 4a has a subagent that never saw the old version read the final text.
What evidence it gives
The source gives the skill's own instructions, README and changelog. It reports 169 linter hits on a specimen file and 0 on clean prose. Some claims cite other repositories and an arXiv paper, which were not opened. No independent test appears.
Interests and limits
The skill is the author's own product, described by its author. Its claims about catching tics come from the author's tests. The read covered the release notes, SKILL.md and a few reference files. It did not cover the scripts, tests, register.md, corpus.md, preservation.md or most model profiles. The scripts were not run. Read in part: the release notes, SKILL.md, README, four references and the changelog head.