Anti-slop skill: audit and fix AI slop in your agent's writing
Most anti-slop prompts are ban lists, and agents follow them loosely. This free skill makes the agent count: ten checks with hard caps, fix only what fails, re-count, and report a one-line tally you can read in a PR.
Download SKILL.md Try the same checks in your browser
Install
Claude Code
mkdir -p ~/.claude/skills/anti-slop curl -o ~/.claude/skills/anti-slop/SKILL.md https://gounslop.com/anti-slop-skill/SKILL.md
Codex
mkdir -p ~/.codex/skills/anti-slop curl -o ~/.codex/skills/anti-slop/SKILL.md https://gounslop.com/anti-slop-skill/SKILL.md
Cursor
mkdir -p .cursor/rules curl -o .cursor/rules/anti-slop.mdc https://gounslop.com/anti-slop-skill/SKILL.md
Then: "anti-slop docs/README.md", or "write the PR description and run anti-slop before you finish".
What it catches
| Check | Cap |
|---|---|
| Throat-clearing openers | 0 |
| Negative parallelisms ("it's not X, it's Y") | 0 |
| "Rather than" / "should be treated as" disclaimers | 1 per 500 words |
| AI vocabulary (crucial, leverage, showcase, multifaceted…) | 1 per 300 words |
| "serves as" / "boasts" for is/has | 0 |
| Rhythm lists of three | 1 per piece |
| Em dashes | 1 per 300 words |
| Bold labels, emoji bullets in short prose | 0 |
| Chat residue ("I hope this helps") | 0 |
| Summary closers | 0 |
It also has rules for commit messages, PR descriptions and code comments, where agents write the most unread slop. The caps are based on our measurements of 11 OpenAI models and Wikipedia's guide.
The skill
---
name: anti-slop
description: Audit-then-fix pass for AI slop in any prose an agent writes or edits (docs, READMEs, PR descriptions, commit messages, emails, posts). Use before returning or committing any text, or when asked to "anti-slop", "unslop", "de-AI" or tighten writing.
---
# Anti-slop
Run two passes. Pass 1 audits and counts. Pass 2 fixes only what failed. Never invent facts, never change code, names, numbers, quotes or links.
## Pass 1: audit (count, don't fix yet)
Count each item in the draft and compare with the cap.
| # | Check | Cap |
|---|-------|-----|
| 1 | Throat-clearing openers ("It's worth noting", "In today's...", "Let's dive in") | 0 |
| 2 | Negative parallelisms ("it's not X, it's Y", "not just X but Y", "X isn't A. It's B.") | 0 |
| 3 | "Y rather than X" / "should be treated as" disclaimers | 1 per 500 words |
| 4 | AI vocabulary: crucial, delve, pivotal, robust, seamless, leverage, foster, showcase, underscore, multifaceted, vibrant, tapestry, testament, landscape (abstract) | 1 per 300 words |
| 5 | "serves as / stands as / boasts" where "is / has" works | 0 |
| 6 | Lists of exactly three used for rhythm | 1 per piece |
| 7 | Em dashes | 1 per 300 words |
| 8 | Bold run-in labels, emoji bullets, headings in short prose | 0 unless the format needs them |
| 9 | Chat residue: "I hope this helps", "Let me know if", "If you'd like, I can" | 0 |
| 10 | Summary closer that repeats the piece ("In short", "Overall", "Ultimately") | 0 |
Write the audit as a one-line tally, e.g. `audit: 2 negations, 3 vocab (cap 1), 1 em dash -> fix 1, 2, 4`.
## Pass 2: fix
- Delete before you rewrite. Most slop is removable without replacement.
- Negations: keep the true half, state it directly.
- Vocabulary: the plain word (crucial -> important or cut; leverage -> use; showcase -> show).
- Lists of three: keep the items that are true and needed, any number.
- Em dashes: comma, period, colon or parentheses.
- Keep the author's register. Don't add slang, typos or jokes to sound human.
## Pass 3: re-count
Re-run the table. If anything is still over its cap, fix once more, then stop. Report the final tally in one line.
## Code-adjacent text
- Commit messages: imperative subject under 72 chars, no "This commit...", no emoji.
- PR descriptions: what changed, why, how to test. No "This PR aims to enhance".
- Code comments: say why, not what. Delete comments that restate the code.
Free online checker with the same categories: https://gounslop.com/ai-slop-detector
Anti-slop skills compared
| Skill | Approach | Good for |
|---|---|---|
| hardikpandya/stop-slop | Banned phrases and structures in reference files, 5×10-point rubric | Essays and posts; the most popular option. Details |
| blader/humanizer | Wikipedia's Signs of AI writing as instructions | General prose |
| adewale/anti-slop-writing | Instruction-only skill with references, many clients documented | Reviewing and drafting prose |
| aixarizzo/kill-slop | Audits against your own voice samples; won't grade without them | Writers with an established voice |
| rand/cc-polymath anti-slop | Skill plus Python detect/clean scripts; text, code and design | Teams who want scripts |
| gounslop humanizer | Wikipedia guide plus 2026 GPT-6 tells | Rewriting to read naturally |
| gounslop anti-slop (this page) | Count → fix → re-count with numeric caps, one-line tally | Agent pipelines, docs, PRs, commit messages |
Pick by workflow, not by name. All of these are free. Read any community skill before you install it: they're instructions your agent will follow.
For text you didn't write with an agent
Paste it into the unslop tool: it rewrites in your tone, keeps facts and numbers, and shows a before/after slop 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 an anti-slop skill?
An Agent Skill (a SKILL.md file) that makes an AI agent like Claude Code, Codex or Cursor check its own writing for AI slop (filler, stock vocabulary, contrast frames, chat residue) and fix it before handing text back.
How is this anti-slop skill different from humanizer or stop-slop?
It's built as an audit with numbers: ten checks, each with a cap (0 negative parallelisms, 1 AI-vocabulary word per 300 words, 1 em dash per 300 words, and so on). The agent counts, fixes only what failed, re-counts, and reports a one-line tally. That makes it predictable in pipelines and code review. Humanizer is a style guide based on Wikipedia's list; stop-slop is a ban list with a 50-point rubric.
How do I install it in Claude Code?
mkdir -p ~/.claude/skills/anti-slop && curl -o ~/.claude/skills/anti-slop/SKILL.md https://gounslop.com/anti-slop-skill/SKILL.md, then restart Claude Code. Use ./.claude/skills/ instead to scope it to one repo.
Does it work in Codex and Cursor?
Yes. Codex reads skills from ~/.codex/skills/ (or .agents/skills/ in a repo). In Cursor, save it as .cursor/rules/anti-slop.mdc. Any agent that takes a markdown system prompt can use the body as-is.
Does it cover code?
Only the text around code: commit messages, PR descriptions and comments. For refactoring AI-written code itself, see the deslop skill page.