De-slop campaign: purge em dashes and AI tells from this repo's own corpus via /ai-slop:audit fix

[kyle-sexton]

The first dogfood run of the ai-slop plugin (branch claude/ai-content-detection-plugins-hlbcfo) measured the backlog this issue tracks:

• 8,058 findings across the 296 instruction-surface files (every SKILL.md, plugin READMEs, AGENTS.md, root README), after vendored upstream material was config-excluded. Nearly all are rule-em-dash (zero-tolerance policy, decided at plan approval).
• ~32,000 findings repo-wide over 1,161 tracked markdown files (pre-exclusion measurement; the post-exclusion number will be somewhat lower).

This is deliberate follow-up work, out of scope for the plugin-introduction PR:

• Purge em dashes (and other confirmed tells) from instruction surfaces first — highest-impact files, worst-first — via /ai-slop:audit fix, which runs the semantic-diff revert guard per file.
• Continue through the remaining corpus incrementally (a few files or one directory per session; a tidy-style lane is a good fit). One big-bang rewrite is explicitly not the plan: every rewrite is prose judgment that needs the guard, and reviewability dies at 8k-line diffs.
• Add the instruction-layer anti-contamination note (e.g. a .claude/rules line): vendored docs under plugins/*/skills/*/vendor/** are reference content, not style guides — do not imitate their formatting (em dashes included). This is the mitigation for vendor material being excluded from the audit rather than rewritten.
• When the purge completes, consider tightening this repo's .claude/ai-slop.json (e.g. stricter rule-of-three threshold, per the maintainer's rule-of-one preference) so regressions surface immediately.

Measurement commands, config, and the per-file close-out loop are documented in the plugin (plugins/ai-slop/) and docs/topics/ai-content-detection/PLAN.md on the branch above.

[claude]

---

Generated by Claude Code

[kyle-sexton]