An editor that treats AI clichés as compiler errors

What it does

This project packages a set of editing rules as a skill for AI coding assistants such as Claude Code and Codex. Feed it a draft and it strips out common LLM tics—binary contrasts, throat-clearing openers, faux-insight setups, and a couple dozen other patterns—while enforcing active voice, concrete numbers, and clearer sentence structure. It returns the edited draft plus a short summary of what changed.

The interesting bit

Rather than vaguely complaining about “AI slop,” the author has reverse-engineered the actual formulaic patterns—colon reveals, weasel attribution, dramatic fragmentation—and turned them into a detectable rule set. The skill even runs its own edits against an internal eval.md pass/fail checklist, which is a tidy bit of recursive quality control.

Key highlights

• Detects 20+ specific slop patterns, from “fake-strong verbs” to “synonym cycling”
• Enforces classic writing discipline: lead with the point, active voice, concrete numbers over abstractions
• Self-checks edits against a built-in eval.md criteria list before finishing
• Designed to make “minimum effective edits” rather than full rewrites
• Runs inside existing AI assistant workflows, not as a separate application

Caveats

• Requires an AI assistant environment (Claude Code, Codex, or similar); it is not a standalone program
• The README does not explain how the eval.md self-check is implemented or how often it catches its own mistakes

Verdict

Anyone who uses LLMs to draft prose but wants to sound less like a chatbot should find this useful. Purely human writers can safely ignore it.