Removes the wording, phrases, and sentences that mark text as AI generated, and replaces them with good human English.
by Mark Stent
Most "de-slop" tools run a blocklist and stop. You get text that is technically clean but flat and voiceless, which is its own kind of tell.
slop-chips runs two passes:
- Strip. Detect every AI tell: em dash overuse, significance inflation, promotional adjectives, rule-of-three, copula avoidance, chatbot artifacts, filler, and the rest.
- Rewrite. Rebuild each sentence toward a distilled style guide of how good non-fiction writers actually write. Concrete, varied, human.
Most tools stop after the first pass. slop-chips does the rewrite too, using a style guide drawn from how writers like Malcolm Gladwell and Stephen Fry actually build a sentence.
| You want to | Ask for | You get |
|---|---|---|
| Fix the writing | "clean this up", "humanize", "de-slop" | The rewritten text plus a short note on what changed |
| Just assess it | "score this", "how much slop", "don't rewrite" | A slop score (0 to 100) and a findings report, no edits |
By default it proposes the rewrite so you can accept it. Say "just fix it" and it edits the file in place.
slop-chips is a plain Agent Skills skill: one folder with a SKILL.md. It works in any harness that reads that standard, including Claude Code, Codex, opencode, and Pi. Installing is always the same idea: clone the repo into that harness's skills directory.
Pick your harness and run one line. This installs it globally (available in every project).
Claude Code
git clone https://github.com/markstent/slop-chips.git ~/.claude/skills/slop-chipsCodex
git clone https://github.com/markstent/slop-chips.git ~/.codex/skills/slop-chipsopencode
git clone https://github.com/markstent/slop-chips.git ~/.config/opencode/skills/slop-chipsPi
git clone https://github.com/markstent/slop-chips.git ~/.pi/agent/skills/slop-chipsCodex, opencode, and Pi all also read a shared location, ~/.agents/skills/. Clone there once and those three pick it up together:
git clone https://github.com/markstent/slop-chips.git ~/.agents/skills/slop-chips(Claude Code reads only ~/.claude/skills/, so it always needs its own clone.)
To ship the skill inside a repo so your whole team gets it, clone into the project instead of your home directory. Use the folder your harness expects: .claude/skills/ (Claude Code), .opencode/skills/ (opencode), .pi/skills/ (Pi), or the shared .agents/skills/ (Codex, opencode, Pi).
git clone https://github.com/markstent/slop-chips.git .claude/skills/slop-chipscd <the folder you cloned into>/slop-chips && git pullAfter installing, restart your harness (or start a new session) so it picks up the new skill.
Two ways to trigger it.
1. Just ask in natural language. The skill activates on its own when your request matches:
Humanize the intro section of README.md
De-slop this paragraph: <paste text>
Score this for AI slop, don't rewrite it: <paste text>
Clean up the wording in docs/guide.md and just fix it in place
2. Call it directly with a slash command. In harnesses that support slash-invoked skills (such as Claude Code), type /slop-chips to run it on demand:
/slop-chips humanize this: <paste text>
/slop-chips score this, don't rewrite: <paste text>
Both routes run the same skill. Use the slash command when you want to be sure it fires, or when the text is ambiguous.
slop-chips overlaps in purpose with skills like humanizer, stop-slop, and no-ai-slop. They do not clash on disk, since each lives in its own folder. The only question is which one runs when you ask in plain language, because their triggers are similar and the harness picks whichever it judges the best match.
To remove the guesswork, name the skill you want:
/slop-chips humanize this: <paste text>
use slop-chips to score this: <paste text>
The slash command and naming the skill both force slop-chips to run regardless of what else is installed. If you would rather not think about it, keep just one de-slop skill enabled.
- Touch code logic. Prose only. It cleans comments and docstrings only when asked, and leaves the code alone.
- Invent facts. It will never add a name, number, or quote to sound concrete. If the specific is not there, the general statement stays.
- Leave em dashes. The output contains zero em dashes and zero en dashes. This is a hard rule, not a preference.
- Flatten your voice. Unusual detail, mixed feelings, first-person choices, and register are preserved. It flags clusters of tells, not every lone dash.
Before:
Nestled at the intersection of design and technology, our groundbreaking new app stands as a testament to seamless innovation. Boasting a vibrant interface and a rich feature set, it empowers users to unlock their full potential. Exciting times lie ahead!
After:
We built an app for keeping notes. It syncs across your phone and laptop, and it opens in under a second. You can search by tag, date, or full text. It costs nothing for the first year.
More in examples/before-after.md.
slop-chips/
├── SKILL.md # entry point: modes, workflow, hard rules
├── references/
│ ├── slop-patterns.md # the detection catalog (strip pass)
│ ├── good-english.md # the distilled style guide (rewrite pass)
│ └── slop-score.md # detect-mode scoring rubric
└── examples/
└── before-after.md # worked transformations
slop-chips stands on the shoulders of people who catalogued these patterns first:
- Wikipedia: Signs of AI writing, maintained by WikiProject AI Cleanup. The detection categories draw heavily on this work.
- stop-slop by Hardik Pandya.
- no-ai-slop by Peter Yang.
Where those projects focus on detecting and removing tells, slop-chips adds the positive half: a style guide it rewrites toward, distilled from real books rather than a banlist.
MIT. See LICENSE.
Built by Mark Stent.