This widens the humanize pack using Wikipedia's signs of AI writing, its companion page on signs of AI use in comments, and a few 2026 vocabulary analyses.
The gap was correspondence. Prose tells were covered: significance inflation, copula avoidance, negative parallelism, superficial analysis. The voice people use when they defend text was not: promises of good conduct, apologies for harm that never gets named, blanket requests for feedback, procedural claims of compliance. Ten new Weir rules cover that, and sixteen existing rules picked up phrases from the comments page and the 2026 lists.
It also adds checks no rule can measure: information density, over-explanation, repeated -ing openers, and transition density. Those live in SKILL.md because they need the whole document.
Two of the points conflict with evidence already in the repo, so I left them out. Grammarly lists "tends to" and "arguably" as AI tells, while Wikipedia's human-signs section has qualifiers like "tends to" appearing more often in human writing. "Perfect grammar" is in the same position: Wikipedia lists it as an ineffective indicator. Adding either would have made the skill delete natural prose. I took the vocabulary hints and the document-level checks only.
mise run test:humanize passes with 53 rules and 253 cases. The README token counts are stale because skill-stats.go needs an API key.