AI Writing Rules

AI writing rules

A Claude Code plugin that stops Claude writing like an AI.

It loads a compact ruleset into every session, ships a skill with 29 researched patterns, and runs a checker after every file write that flags the tells with line numbers and tells Claude to fix them.

Built from Wikipedia's "Signs of AI writing" plus measurement work from SlopDetector and others, turned into rules a model can actually follow.

Install

Restart Claude Code. The rules load on the next session start.

What it does

Loads rules at session start. A SessionStart hook injects the hard rules into context. No em dashes, straight quotes, sentence case headings, no "not just X, it's Y", no three-item-list reflex, no wrap-up endings, plus the vocabulary to avoid. About 270 words, so it costs little per session.

Checks what Claude writes. A PostToolUse hook runs after every Write and Edit on .md, .mdx, .markdown, .txt, and .rst. It scans for 17 regex-detectable patterns plus three statistical ones, and when it finds something it reports the line numbers to Claude and asks for a fix before moving on.

Example output:

Gives Claude the detail on demand. The anti-ai-writing skill carries the full ruleset and points at one research file per pattern, so Claude can look up a specific tell rather than loading 29 files into context.

Rewrites on request. /deslop path/to/file.md audits a file, rewrites it, and re-runs the checker to confirm it comes back clean. With no argument it works on the last thing Claude wrote.

What gets checked

Regex rules with line numbers:

em dashes, curly quotes and the ellipsis character, negative parallelism, AI vocabulary, inflated verbs, dodged copulas, puffed significance, vague attribution, empty openers, outline endings, hedge stacks, pseudo-wisdom filler,…