What it does

Stop Slop is a skill file — a bundle of instructions, banned-phrase lists, and structural rules — that you feed to Claude (or another LLM) to make it edit its own output. The goal is stripping the predictable rhythms, throat-clearing openers, and business-jargon crutches that mark text as AI-generated. It scores drafts on a 50-point rubric across five dimensions; anything below 35 gets sent back for revision.

The interesting bit

The project treats LLM slop as a pattern-matching problem and fights it with more pattern-matching — a pleasingly recursive approach. The scoring rubric is the unusual part: most “humanize” tools are vibes-based, but this one forces a numeric threshold. The README is also admirably direct about what it bans (all adverbs, no em dashes, no Wh- sentence starters), which reads like a cranky copy editor’s revenge.

Key highlights

• Ships as a Claude Code skill with SKILL.md plus three reference files (phrases, structures, before/after examples)
• Bans specific categories: throat-clearing openers, binary contrasts, false agency, passive voice, “lazy extremes”
• 50-point scoring system with five dimensions: Directness, Rhythm, Trust, Authenticity, Density
• MIT licensed; designed for drop-in use across Claude Code, Claude Projects, custom instructions, or raw API calls
• 8.4k stars suggests the itch is widely felt

Caveats

• The README doesn’t show actual before/after examples — you have to dig into references/examples.md
• Some rules (no em dashes? all adverbs?) may produce prose that feels flattened rather than genuinely human
• Effectiveness depends heavily on the base model’s willingness to follow instructions; no benchmark data provided

Verdict

Grab this if you publish AI-assisted writing and worry readers can smell the syntheticity. Skip it if you want creative or voiceful prose — the rules lean toward a narrow, aggressive plainness that may just trade one tell for another.