Deslop skill: which one you want, and how to install it
"Deslop" means two different things in agent land. One kind cleans AI-written code in a branch diff; the other cleans AI-written prose. Here's how to tell them apart, how to install each, and a free skill that does both for a pull request.
The deslop skills people mean
| Skill | Cleans | Install |
|---|---|---|
| Cursor's deslop (cursor/plugins, cursor-team-kit) | Code: extra comments, abnormal defensive checks and try/catch, casts to any, deep nesting, style that doesn't match the file. Works on the diff against main. | npx skills add cursor/plugins --skill deslop |
| ozdemircili/deslop | Prose: AI tells, with a zero-dependency Python linter that scores a draft line by line | git clone into ~/.claude/skills/deslop |
| fayerman-source/deslop | Prose: plain-English rules for legal writing (after Garner) | npx skills add fayerman-source/deslop |
| gounslop deslop (below) | Both halves of a PR: the code diff and the text in it (comments, docstrings, README/CHANGELOG, commit and PR text) | curl, see below |
Free deslop skill for pull requests
Code-only deslop leaves "This PR aims to enhance the overall robustness" in the description. Prose-only deslop doesn't touch the five try/catch blocks. This one does both, scoped to what the branch changed, and never changes behavior.
Claude Code
mkdir -p ~/.claude/skills/deslop curl -o ~/.claude/skills/deslop/SKILL.md https://gounslop.com/deslop-skill/SKILL.md
Already have Cursor's deslop under that name? Use ~/.claude/skills/deslop-pr/ and change name: to deslop-pr.
Codex
mkdir -p ~/.codex/skills/deslop curl -o ~/.codex/skills/deslop/SKILL.md https://gounslop.com/deslop-skill/SKILL.md
Cursor
mkdir -p .cursor/rules curl -o .cursor/rules/deslop.mdc https://gounslop.com/deslop-skill/SKILL.md
Then: "deslop this branch before I open the PR".
The skill
---
name: deslop
description: Remove AI slop from a branch before review, in both the code diff and the text around it (comments, docstrings, commit messages, PR description, README/CHANGELOG edits). Use when asked to "deslop", clean up AI-written changes, or prepare a PR.
---
# Deslop (code + repo text)
Scope: only what this branch changed. Never change behavior. Keep edits minimal.
## 1. Find the diff
- Base: origin/main, else origin/master, else main/master.
- `git diff --name-only <base>...HEAD` then `git diff <base>...HEAD`.
## 2. Code slop (behavior-preserving)
- Comments that restate the code, narrate steps ("// Now we loop over items"), or address the reader ("// Note: this is important"). Delete them. Keep comments that explain why.
- Defensive checks, try/catch and null guards that the rest of the file doesn't use for the same trusted paths. Remove them unless a caller can really pass bad input.
- One-use wrappers, helper functions and abstractions added "for flexibility". Inline them.
- Type escapes added to silence errors (`as any`, `# type: ignore`, `!`). Fix the type or leave a TODO with the reason.
- Deep nesting that an early return removes.
- Logging added for debugging and left in.
## 3. Text slop in the diff
- Docstrings and README lines: no "seamlessly", "robust", "leverage", "comprehensive", "This function is designed to...". Say what it does in one sentence.
- No "It's not X, it's Y", no emoji bullets, no bold run-in labels in docs that don't use them elsewhere.
- CHANGELOG: one line per change, past tense, no marketing.
## 4. Commit and PR text
- Commit subject: imperative, under 72 characters, no emoji, no "This commit...".
- PR description: What changed / Why / How to test. No "This PR aims to enhance the overall robustness".
## 5. Report
One to three sentences: what you removed, and anything you left on purpose and why.
Check any prose output for free: https://gounslop.com/ai-slop-detector
Deslop vs anti-slop vs humanizer
- Deslop: after the fact, on a branch. Code first, text second.
- Anti-slop: any prose an agent writes, with counted caps and a re-count.
- Humanizer: rewriting prose so it reads naturally, based on Wikipedia's guide and 2026 tells.
- Stop slop: a popular ban-list skill for essays and posts.
Not to be confused with desloppify, a separate open-source CLI that scores codebase quality 0–100 and runs a fix loop.
Check the text part yourself
Paste a PR description, README section or release note into the AI slop detector to see what's left. For longer docs, the unslop tool rewrites with a before/after score.
The skill and checklists are free. For finished text from ChatGPT or Claude, the unslop tool rewrites with a slop score. After the 3 free daily rewrites: $5 for 50 or Pro $12/month.
FAQ
What is a deslop skill?
An Agent Skill that cleans up "slop" an AI left behind. The word means two things: in code, it removes AI-style clutter from a branch diff (extra comments, needless try/catch, wrappers, type escapes); in prose, it removes the tells of AI writing. Check which one a skill does before installing it.
Where does the Cursor deslop skill come from?
Cursor's public plugins repo (github.com/cursor/plugins) includes a deslop skill in its cursor-team-kit. It reads the branch diff against main and removes AI-generated code slop. A common install is: npx skills add cursor/plugins --skill deslop.
Is there a deslop skill for writing, not code?
Yes. ozdemircili/deslop on GitHub is a prose skill with a zero-dependency Python linter, and fayerman-source/deslop applies plain-English legal writing rules. For general prose, our humanizer and anti-slop skills cover the same ground.
What does gounslop's deslop skill do differently?
It handles both halves of a pull request: the code diff (behavior-preserving cleanup) and the text in it (comments, docstrings, README and CHANGELOG lines, commit messages and the PR description). Most deslop skills do one or the other.
How do I install it in Claude Code?
mkdir -p ~/.claude/skills/deslop && curl -o ~/.claude/skills/deslop/SKILL.md https://gounslop.com/deslop-skill/SKILL.md. If you already installed Cursor's deslop under the same name, put ours at ~/.claude/skills/deslop-pr/ instead.
What does desloppify mean?
Desloppify is a separate open-source CLI (by Peter O'Malley) that scans a codebase, scores its quality 0–100, and gives an agent a fix loop. It's a tool, not a SKILL.md, and it's for code.