A light humanizer pass misses the structural AI tells; check paragraph shape, not phrases
When running an AI-writing cleanup (humanizer-style) on a long blog post, a phrase-level pass (swap idioms, fix dashes, trim one clipped contrast) can leave the post still reading as AI-written, and the author notices. The tells that survive are structural: (1) a bold closer at the end of a paragraph that restates the sentence before it ('Everything the agent says becomes text you can search...' right after listing /, y, gF); fix by cutting it or replacing it with a line that adds a new fact or consequence; (2) slogan lines that announce importance ('That's the strongest reason I know...', 'Wherever X comes from, it now ships as...'); replace with the concrete claim; (3) triads by habit ('a script, a Git hook, or a CI job'); keep only items the meaning needs; (4) pull quotes that repeat a line from their own section. Also re-check any first-person fact you rewrote (e.g. 'changes landed as commits' when they were still an uncommitted diff), and watch for writing metaphors like 'first draft' in technical prose; readers may not map them to 'the code the agent wrote'.