What problem does it solve? Prose drafted or edited by AI models tends to fall into a recognizable default register: em dashes, "it's not just X, it's Y" cadences, assistant boilerplate, diction like "delve" and "tapestry", listicle scaffolding, and tidy conclusions. Readers flag these cues as machine-written, and naive fixes often over-correct into an equally detectable "trying not to sound like AI" voice. This Skill removes the cited tells and forces a deliberate, defensible voice instead of the model's median. ## Core Features & Use Cases - Deterministic scanner: Run scripts/unslop_text_scan.py on .md, .markdown, .mdx, .txt, .rst, or .html files to get per-line findings with severity, data share, suggested fixes, a slop score, and density per 1,000 words; the exit code equals the high-severity count so CI can gate on it. - Data-grounded tell catalog: references/tells.md ranks tells by how often real readers cite them, based on an analysis of 89,239 Reddit posts, and separates surface lexical tics from structural tells like uniform rhythm and sycophancy that must be checked by ear. - Build and Audit modes: Build mode pins one of four registers (casual, conversational-professional, expository, formal) plus a speaker, claim, and shape before drafting; Audit mode scans existing prose and fixes in priority order while honoring unslop-ignore lines for intentional choices. - Use Case: A user pastes a blog draft that "sounds like ChatGPT." Run the scanner for the lexical layer, then read the draft aloud against the structural catalog, replace flagged phrases with plain verbs, and report a verdict with the top three changes. ## Quick Start Ask the assistant to audit your draft for AI-writing tells, for example: "Run the unslop scan on my post.md and rewrite anything that reads as AI-generated in my own voice."
