feat(skills): add humanize, the ai-tells catalogue and Vale style; v0.4.0

[sebastian-iancu]

Summary

Adds /humanize, which removes AI writing patterns from existing prose, and makes the house style em-dash free. Prepares v0.4.0.

Three commits:

1. feat(skills): the humanize skill, the catalogue and the Vale style.
   • skills/humanize/: picks an edit contract (line edit by default), runs the ai-tells Vale style, marks tells by strength, restates without adding or dropping facts, and reports. --report-only names patterns and never judges authorship.
   • references/ai-tells.md: the canonical catalogue. It covers leftovers, staging, inflation, rhythm and formatting, a strong versus weak-alone ranking, the em-dash rule, a when-not-to-act list, and why tells describe writing, not writers. It is written from scratch and sourced to Wikipedia's "Signs of AI writing" and Liang et al. (2023).
   • references/vale-styles/ai-tells/: seven rules.
     • Errors: chatbot-residue, tool-markup.
     • Warnings: signposting, stock-vocabulary, vague-attribution.
     • Suggestions: negation-contrast, em-dash. A person's deliberate dash never fails CI.
   • One-line hooks in the router, copy-editing, technical-writing, marketing-copy and prose-reviewer. /docs-lint-setup copies the style into a repo.
2. style: every em dash removed from the plugin's own prose. Each replacement was chosen per sentence: a colon, comma, semicolon, parentheses or a new sentence. Punctuation only; no rule changes strength.
3. chore(release): v0.4.0. Both manifests are bumped and the changelog is folded. The tag is cut on main after merge.

The em-dash rule (references/ai-tells.md §6, style-guide.md §3):

• The assistant writes no em dashes.
• An existing dash a person wrote stays, except in a humanize pass, and even there only where other tells share its paragraph.
• Vale reports dashes as a suggestion. An author can silence a passage with <!-- vale ai-tells.em-dash = NO -->.

Testing done

• Vale 3.18.0 fixtures:
  • A sloppy sample fires every rule.
  • A clean sample stays silent, including code, ranges like 2020–2026 and "not only … but also".
  • Fifteen false positives the skill review predicted were confirmed, then fixed. Examples: "as an AI pair programmer", "If it's not set, it is created", Apache Tapestry, "teams report incidents".
  • Typographic apostrophes now match.
  • vale sync leaves styles/ai-tells/ in place.
  • A missing style directory fails with E100 [loadStyles].
• Behaviour test: a fresh agent ran the skill on a trap page. It cut an unsourced "studies show" without inventing a replacement. It kept a person's dash, a dash inside a code block, a legitimate "sync is not a backup" warning and a figure that came with its date. It reported the recap section instead of deleting it.
• Skill review (plugin-dev:skill-reviewer): 1 critical and 8 major findings, all addressed.

Checklist

• ./scripts/validate.sh passes (or claude plugin validate . if Python is unavailable)
• claude plugin validate . passes. Its two warnings about unquoted ${CLAUDE_PLUGIN_ROOT} in hooks.json predate this PR.
• Skill / agent triggering tested locally (see docs/testing.md). The skill was run by an agent following it directly; auto-triggering with --plugin-dir has not been tested yet.
• Cursor install tested locally when manifest, rule, or hook content changed (see docs/install.md). Not tested; the Cursor rule and manifest changed.
• Both manifests kept in sync when metadata changed: .claude-plugin/plugin.json and .cursor-plugin/plugin.json
• Rules cite references/ rather than restating them; no rule duplicated across components
• No unsourced claim added: no invented statistics, testimonials, customer names, awards, urgency, or superlatives, in the components or in this repo's own docs (references/claims-and-evidence.md)
• Inventories and counts in README/AGENTS.md verified against the tree (8 skills, 7 worker skills, 5 canonical references)
• New or renamed component wired into the router and hooks/session-start.sh (the validator enforces this)
• Version bumped and CHANGELOG.md updated (if component content changed); see docs/versioning.md

After merge: tag v0.4.0 on main, publish a GitHub release titled v0.4.0, then repin docs-editing in Cadasto/plugin-marketplace.

🤖 Generated with Claude Code