Stop-Slop is a skill that catches AI writing red-handed
You know the tells by now. The throat-clearing opener. The not-just-X-but-Y construction. The em-dash everywhere. The tidy three-part list. Every single adverb. Once you can see AI writing you can't unsee it, and it's everywhere. Stop-Slop, trending on GitHub this week, is a skill that teaches the model to catch itself.
It's the same idea as Taste-Skill but pointed at prose instead of pixels. You hand Claude or any model the skill and it learns to flag its own patterns, banned phrases, binary contrasts, passive voice, vague declaratives, meta-commentary, then rewrites. It scores the result on five axes: directness, rhythm, trust, authenticity, density. It ships with before-and-after examples so the model learns what less slop actually looks like.
What makes it interesting is the meta-move. Instead of a separate detector tool that scans output after the fact, it makes the model police itself from the inside, baked straight into the system prompt. The thing producing the slop becomes the thing trained to remove it.
Stop-Slop and Taste-Skill both landing on trending in the same week is not a coincidence. The first wave of skills was about teaching agents to do more. This wave is about teaching them not to be generic. When everyone runs the same model, the only thing left to compete on is taste, and now you can install it.
Repo: github.com/hardikpandya/stop-slop
← Back to all articles
It's the same idea as Taste-Skill but pointed at prose instead of pixels. You hand Claude or any model the skill and it learns to flag its own patterns, banned phrases, binary contrasts, passive voice, vague declaratives, meta-commentary, then rewrites. It scores the result on five axes: directness, rhythm, trust, authenticity, density. It ships with before-and-after examples so the model learns what less slop actually looks like.
What makes it interesting is the meta-move. Instead of a separate detector tool that scans output after the fact, it makes the model police itself from the inside, baked straight into the system prompt. The thing producing the slop becomes the thing trained to remove it.
Stop-Slop and Taste-Skill both landing on trending in the same week is not a coincidence. The first wave of skills was about teaching agents to do more. This wave is about teaching them not to be generic. When everyone runs the same model, the only thing left to compete on is taste, and now you can install it.
Repo: github.com/hardikpandya/stop-slop
Comments