Catch AI patterns that appear across a whole draft

[vivekkartha]

What this improves

Humanizer now catches several habits that only become obvious across sentences or paragraphs: split correction formulas, three-example cadence, empty bridge lines, and repeated mini-conclusions. It also catches source names used only to borrow credibility instead of telling the reader what the source adds.

The second pass now checks the shape of the full draft, so it can rewrite repeated structures instead of patching one phrase at a time.

False-positive safeguards

The new guidance keeps real distinctions, meaningful transitions, necessary examples, deliberate short paragraphs, and citations that support a fact or useful finding. It also keeps the existing rule against adding claims or sources that were not in the input.

Validation

• python3 scripts/validate-package.py
• npx skills add . --list
• npx -y @anthropic-ai/claude-code plugin validate .

Note: In essays, a prestige source in the sentence is an empty credibility signal even when the source is real. State the claim; move the source to references. validate-package.py now rejects example Afters that miss §10 / §28 / §31 / §36.

[blader]

Closing, with thanks for the core insight: these patterns also appear at paragraph scale, as a contrast split across sentences, three parallel examples, or the same one-line conclusion after several sections. That is worth one sentence in step 1. The PR as a whole adds a validator that pins exact example sentences, a new pattern that moves real citations out of the sentence, and edits to old release notes, so it cannot land as is.