slop-score

"This is not AI-detection, but slop detection." — scores 25.32/100 [SLOP]. The tagline fails its own test.

AI-detection asks who wrote a sentence; it's unreliable, adversarial, and increasingly beside the point. Slop detection asks the question that actually matters — does the writing read like slop? — and that one can be counted, with receipts. No authorship verdict, no vibes: just the exact overused words, trigrams, and "not X, but Y" constructions, tallied and scored.

An agent skill (Claude Code, Cursor, Codex, or any SKILL.md-compatible agent) that runs Sam Paech's eqbench Slop Score heuristic against a piece of writing. Faithful port of https://eqbench.com/slop-score.html — composite of slop words (60%), "not X, but Y" contrast patterns (25%), and slop trigrams (15%), min-max normalized against the model leaderboard, scaled 0–100. > 25 = SLOP. It flags, never rewrites — you stay the editor.

Install

npx skills add closestfriend/slop-score-skill     # Claude Code, Cursor, Codex, etc.

Or clone straight into your agent's skills directory:

git clone https://github.com/closestfriend/slop-score-skill ~/.claude/skills/slop-score

Zero-install scoring out of the box — Stage 2 is the only thing that needs npm install (see below).

Usage

node scripts/slop-score.mjs --file path/to/draft.md        # human readout
node scripts/slop-score.mjs --file path/to/draft.md --json # machine JSON
node scripts/slop-score.mjs --text "some short text"
cat draft.md | node scripts/slop-score.mjs --stdin

Stage 2 (optional, fuller fidelity)

Out of the box it scores slop words + trigrams + Stage-1 surface contrast patterns with zero install. The 35 Stage-2 POS contrast patterns need wink-pos-tagger:

npm install        # local to this folder; unlocks Stage 2

The script auto-detects it; if absent, it scores without Stage 2 and says so. node_modules/ is gitignored so shared clones stay zero-install.

Regenerating normalization ranges

If the upstream leaderboard changes:

node scripts/build-normalization.mjs path/to/leaderboard_results.json

Tests

node --test

Credit & License

MIT — see LICENSE. Heuristic and slop lists by Sam Paech (eqbench.com / slop-forensics, MIT).