Anti-AI-Slop Writing Skill
github.com/jalaalrd/anti-ai-slop-writing- Category
- Developer Tools
- Rank
- No. 1906Tools index
Previous survey · No. 1914 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Productivity & Collaboration
- Interfaces
- Agent Skill / Plugin
- Builder
- jalaalrd
- GitHub
- 482 stars
- Date
About
A SKILL.md-based instruction set that constrains AI text generation to eliminate statistically detectable AI writing patterns—banned words, phrases, sentence openers, structural tics, and formatting leaks—drawing on Carnegie Mellon research and large-scale post analysis. It installs into Claude Code, Cursor, Codex, Gemini CLI and other agents to make AI-generated text read as human-written.
What it does
A plain prompt, two Markdown documents and nothing else, that tells a coding or chat assistant how to write prose people won't dismiss as machine output. One document holds the rules: sentence rhythm, punctuation ceilings, honesty about numbers and quotes, and an eleven-question checklist the model runs on its draft before replying. The second is a long denylist of words, stock phrases and openers, which the rules tell the model to pull in before it starts writing. There is no code; the assistant reads the text and is expected to obey it.
Why it's ranked here
Its rules come with numbers a reader can check against the output: one em dash per 500 words, one exclamation mark per 1,000, at most one ellipsis per piece, zero to two hashtags. It also treats fabrication as a writing failure, forbidding invented statistics, quotes and anecdotes, which most style prompts skip. MIT licensed, a two-file install and a paste-into-chat fallback make it cheap to try. The ceiling is that it is only a prompt: nothing verifies compliance, and several rules rest on research the files name but never link.
What's good
The closing checklist turns style advice into questions a model can answer yes or no, such as whether three consecutive sentences share a length or whether any specific was invented. The banned list is split from the core rules and loaded when needed, so the main instructions stay short. Rules explain the replacement, not just the ban: swap em dashes for commas, colons or parentheses; turn a sentence already fine as prose into prose instead of bullets. A voice section gives a default register for when the target writer is unknown.
Tradeoffs
Nothing enforces any of it. There is no linter, script or test, so results depend entirely on how faithfully the model follows a long prompt. Some advice pulls against itself: it urges grounding text in moments like "last Tuesday" or "at 2am" while forbidding invented anecdotes. The denylist is broad, banning ordinary words like "crucial" and "highlighting" and all passive constructions, which can make technical writing stilted. The trigger description fires on almost any writing request, including READMEs. The one-command plugin install points at a different repository from this one.
How to use it well
Best for people drafting posts, emails and landing copy with an assistant who want a consistent house voice with fewer stock phrases. Install it as a skill so it loads on writing requests, or paste it into a chat as a standing instruction. Read the output against the checklist yourself, since the model is told to apply the rules silently. For documentation or specs, where passive voice and precise terms are sometimes right, relax the passive and vocabulary bans. It does not fact-check claims or detect machine text; it only shapes what gets written.
Technical notes+
The repository is three Markdown files. skills/anti-ai-slop-writing/SKILL.md carries YAML frontmatter (name, description) whose description lists trigger phrases such as "write", "draft", "rewrite" and "anti-slop", followed by sections on structure, punctuation, accuracy, formatting, voice calibration and an 11-step self-check; it instructs the model to load the companion banned-words file before writing. skills/anti-ai-slop-writing/references/banned-words.md holds a comma-separated vocabulary list, 36 banned phrase patterns, 16 banned openers, per-model first-word tells for five model families, and era-grouped vocabulary lists marked as context. README.md documents a plugin-marketplace install that references a separate founder-toolkit repository, a manual copy into the skills folder, and a slash-command invocation. No executable code, tests or CI configuration appear in the tree.
Observed
- License
- MIT
- Language
- Markdown only, no executable code
- Interface
- Agent skill with slash-command invocation
- Install surface
- Plugin marketplace (via a separate repository), manual copy into skills folder, or paste into any chat
- Files
- skills/anti-ai-slop-writing/SKILL.md and skills/anti-ai-slop-writing/references/banned-words.md
- Tests
- No test directory in the repository tree
Read from README.md, skills/anti-ai-slop-writing/SKILL.md, skills/anti-ai-slop-writing/references/banned-words.md.
What it can do
Filter AI-generated text to remove statistically detectable AI writing patterns (banned words, phrases, sentence openers, structural tics, formatting leaks)
AI-generated text → Text that reads as human-written
Install as a skill instruction set into AI coding agents (Claude Code, Cursor, Codex, Gemini CLI)
SKILL.md file → Configured agent with anti-AI-slop writing constraints
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.