fuck-slop

This skill is a de-slopping pass for any text: it detects and removes the statistical fingerprints of AI-generated writing and rewrites the text into its intended register. It targets the tells that emerge from how language models generate prose, above all negative parallelism (the "not X but Y" family), plus em-dash abuse, rule-of-three constructions, false ranges, puffery vocabulary, uniform sentence cadence, and hedged both-sidesing. The premise is that these patterns are invisible on self-review, so detection must be mechanical (regex against a fixed catalog) rather than intuitive.

It runs a loop: fix the target (genre, venue, audience, stance, constraints), perform a mechanical scan using runnable grep patterns from a tell catalog plus structural checks for cadence and formatting, report findings as a short diagnosis table, then rewrite by meaning rather than by paraphrasing the pattern. A three-way triage handles negative parallelism (delete strawman negations, earn genuine contrasts with specifics, or cut empty sentences entirely), and banned escape hatches prevent reintroducing the same move in disguise. Rewrites are re-scanned until clean, followed by a register check. A voices reference defines what good writing means per genre (academic paper, tweet, Reddit, LinkedIn, email, blog, marketing, docs, cover letters) and which tells are fatal in each; a tells reference holds the detection catalog.

It targets writers, editors, and anyone publishing agent-drafted prose who wants text that reads like a specific person wrote it. It runs read-only text-analysis operations (grep against a draft file) with no risky behavior.