slop-chips
An agent skill that finds the wording that marks text as AI-written, then rewrites it into clear, human English using a style guide drawn from real books.
Details
- Started
- 8 Sep 2026
- Type
- Agent Skill (a folder with a SKILL.md)
- Runs in
- Claude Code, Codex, opencode, Pi
- Modes
- Rewrite, or score from 0 to 100
- Output
- Zero em dashes, by rule
- Licence
- MIT
- Last commit
- 8 Sep 2026
ConjectureRemoving the tells is not enough: text only reads as human once it is rewritten towards how good writers actually write.
FindingStill open.
The write-up
Most “de-slop” tools run a blocklist and stop. What you get back is technically clean but flat, which is its own kind of tell.
slop-chips does two passes. The first strips the tells: em dash overuse, inflated significance, promotional adjectives, the reflexive list of three, chatbot leftovers and filler. The second rebuilds each sentence towards a style guide distilled from how good non-fiction writers actually write, so the result reads as concrete and varied rather than merely scrubbed.
Two ways to use it
Ask it to clean something up and it proposes a rewrite, with a short note on what changed. Ask it to score something and it leaves the text alone, returning a slop score from 0 to 100 and a list of findings.
What it will not do
- Touch code. It works on prose only, and cleans comments and docstrings only when asked.
- Invent facts. It never adds a name, number or quote to sound concrete. If the specific is not there, the general statement stays.
- Flatten a voice. Unusual detail, first-person choices and mixed feelings are kept. It flags clusters of tells, not every lone dash.
It ships as a plain Agent Skill, so one git clone into the right folder installs it in any harness that reads the standard.