An agent skill that strips AI writing tells (AI-isms) from prose so it reads as human-written.

It loads as an agent skill through the SKILL.md frontmatter, but the two files stand alone and need no framework:

• SKILL.md: the scan, judge, fix, verify procedure and the fix-vs-keep judgment table.
• scan.py: a read-only, dependency-free scanner. It inventories AI-isms as file:line:category:snippet and exits non-zero while any em or en dash remains, so it can gate a verify loop.

python3 scan.py <files-or-dirs>

Add --format json for machine output, --ext md,txt to widen the file types on a directory walk, or --help for all options.

The scanner flags em and en dashes everywhere, and flags word-category tells (hype words, throat-clearing openers, marketing nouns, and the like) only outside code. Em and en dashes are the near-certain tells; the rest need a human call, so read each hit in context before you change it. Code, CLI flags, arrows, and text inside quotes are left alone.

The canonical AI-ism list lives in the Writing Style section of the author's personal ~/AGENTS.md and evolves over time. scan.py encodes a portable subset of it.