Stop Slop
A drop-in skill file that makes an LLM grade its own draft on five axes and rewrite it whenever the total falls below 35 out of 50 — the blacklist of AI tells, wired into a loop.
Everyone can now spot the tells: the "Here's the thing," the "Let that sink in," the reflexive rule-of-three, the em-dash on every other line. Wikipedia keeps a page of them; dozens of "300 AI words to avoid" lists circulate. Stop Slop, a MIT-licensed skill file that reached #3 on GitHub's trending list and roughly 13,800 stars in about six weeks, does something slightly more than catalogue them. It hands the model a rubric and makes it grade its own work.
The skill is a folder an agent reads: a core rulebook plus reference files of banned phrases, banned structures, and before/after examples. The rules are blunt — no em dashes, no "So" or "Look," openers, active voice, kill the tripling. Then the loop: score the draft 1 to 10 on Directness, Rhythm, Trust, Authenticity, and Density, and if the five add to less than 35 out of 50, revise and score again. It turns a static list into a revise-until-passing pass, and because it rides the same SKILL.md format every major coding agent now reads, the identical file runs unchanged in Claude Code, Codex, Cursor, and Gemini CLI.
The catch is baked into the mechanism. The skill asks a model to rate its own draft on "Authenticity: sounds human?" — using the very model whose defaults produced the slop. It also amputates a legitimate human device, the em dash, to kill an AI habit, which is why one practitioner titled a rebuttal "AI Can Pry the EM Dash Out of My Cold, Dead Hands." The sharper objection, from a skeptical write-up called "You Can't Install Taste," is that the goal isn't to make the machine write well — it's to make it stop writing like a machine, which is a smaller thing. But that smaller thing is exactly what most people drafting with an LLM actually want, and here it's a file you can drop in today and read in full.
The lenses
The facts
How this connects
Tap a node to open it