This page explains the scoring method. To edit a draft, open the free editor. Scoring stays on your device; requesting an AI rewrite sends your draft to our hosted service.
The Zero Slop checker runs on your machine. It gives a draft a score from 0 to 100 and shows the passages that moved the number. Lower is better. Twenty-five is the tool’s review threshold, not a claim about who wrote the text.
What the score is built from
Four groups of checks build the score: 294 weighted patterns, sentence variety, readability, and formatting. Each finding points to a phrase or measurement you can inspect. The formula is public.
The phrase checks look for stock constructions, a 96-term watchlist, and 26 context-sensitive terms. Words like “robust” and “leverage” are ordinary technical vocabulary, so they only count when a marketing trigger shares the sentence: “elevated write volume” in a runbook stays silent, while “elevate your brand with our seamless platform” does not.
Sentence rhythm measures how much sentence lengths vary. Research can find aggregate differences between human and model-written samples, but one document may break the pattern. Readability checks flag comma pileups, clusters of long words, and sentences past 38 words. Formatting checks count emoji, heavy bolding, hashtag clusters, and frequent dashes.
A curve turns those checks into one number. Its minimum is 9.5 and it moves fastest through the middle. Anatomy of the slop score walks through the derivation and the three public tests it was checked against.
How it avoids obvious false alarms
Every proposed pattern is run against a corpus of twelve documents that must never be flagged: the Gettysburg Address, a Federalist paper, dense technical prose, terse engineering notes, a business memo, human press copy, and an engineer’s email written in second-language English. A pattern that flags any of them is rejected, whatever else it catches.
The Gettysburg Address, for example, uses em dashes and still scores clear. The checker caps punctuation instead of banning it. It also skips short quotations when matching phrases, so discussing a cliche does not count as using it.
Why this is not an AI detector
This checker is not an authorship detector. In the RAID benchmark, detectors weakened under attacks, unfamiliar models, and different generation settings. The slop score describes the writing, not the writer. A human can write slop, and a machine-assisted draft can read clean.
How to check a draft
To check a draft, install the skill once and ask your agent to score it. It works in Codex, Claude Code, Cursor, Gemini CLI, and other assistants that read SKILL.md files. The standalone local scorer makes no network calls. If you use an AI assistant to rewrite the draft, that assistant’s data policies apply. Local scoring does not make a hosted assistant private.
The scorer uses Python’s standard library and makes no network calls. Ask for the detailed view and it returns each flagged span with its reason and location, ordered by impact.
Fixing what the checker flags
The same skill can rewrite the draft and check names, numbers, quotes, and links against the original. In a public replay of our AI Slop test corpus, deliberately obvious drafts moved from a mean of 76.3 to 12.8. See the measured comparison of four open-source editors, or start with what AI slop is. To score a whole directory instead of one draft, run the same check in CI.