What problem does it solve? Authors cannot see where their own documents stall a first-time reader, because rereading supplies missing context from memory. This Skill replaces author review with a cold read by a fresh agent, classifies each stall against eight mechanical criteria, and repairs the document in a bounded loop until findings fall below a threshold. ## Core Features & Use Cases - Cold-read repair loop: A new agent each round reports where it stalled, findings are classified by severity (critical, major, minor, out-of-scope), and only in-scope defects are fixed, looping until the threshold is met or the round cap is reached. - Multi-language correspondence: Language versions of one document are repaired in parallel per language, then compared for claims, heading structure, figures, and term mapping, with structural changes applied to the whole family at once. - Voice preservation: A style probe extracts the document's register, formatting habits, and vocabulary before editing, and a seam test verifies new passages are indistinguishable from existing ones. - Use Case: A repository's Japanese quick-start guide reads like a translation and has drifted from its English counterpart. Run the skill on the file with lang:all to repair both versions, align their claims and structure, and verify each with a final cold read. ## Quick Start Ask the agent to run /hos-humanize-docs on a document path such as docs/quick-start.ja.md, optionally adding lang:all, rounds:5, strict, or dry to control the repair.
