Humanize content
Talk like ../../references/plain-language.md. Adapted from MIT-licensed humanize workflows (see ../../references/ATTRIBUTIONS.md).
Modes
- Rewrite: user pastes a draft → strip AI patterns only.
- Generation: user asks for new copy → apply rules while drafting.
Hard rules
- Never add facts, numbers, quotes, names or citations that were not in the source or brief.
- Keep quotes, stats, dates, URLs and proper nouns verbatim in rewrite mode.
- Do not change meaning. If removing a pattern would change a claim, keep the claim and only rephrase.
- Detect register first (social / email / editorial / formal) and match it.
Web connector mode
Call check_ai_content with the draft, rewrite in chat, then call check_ai_content again. No Citebase fee; no Apify spend.
Workflow
- Lint first when a draft exists:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/ai_content_lint.py" --json(web:check_ai_content). - Rewrite using
../../references/ai-writing-patterns.md(P1–P36 priority table). - Re-lint until density is under threshold or the user accepts remaining hits.
- Deliver the cleaned text. Add a change list only if asked.
- Done / Cost so far (no paid checks) / Next — offer
eeat-auditorcontent-optimizeif they want trust/SEO next.
Out of scope
Code, commit messages, legal boilerplate that must stay formulaic, and any request to "beat" a named commercial AI detector by gaming it.