CLI tool that scans files and websites for words, phrases, and structural patterns statistically overrepresented in LLM output. Based on frequency data from 67 AI models.
$ slopsquid scan docs/ --json
Built on the Antislop paper (Paech et al., 2025) analyzing 67 AI models. Every detection is weighted by how many models overuse that pattern.
45 banlist words, 27 trigram phrases, and 5 structural patterns — each with frequency ratios derived from real model analysis.
Point it at a URL and get a full site report. Respects robots.txt, seeds from sitemap.xml, configurable depth and concurrency.
Every file gets a 0-100 score based on weighted hits per 1000 words. Rated clean, moderate, or heavy for quick triage.
Scans Markdown, HTML (tags stripped), plain text, reStructuredText, AsciiDoc, and XML. Handles text extraction automatically.
Every command supports --json for integration
with other tools, CI pipelines, or custom reporting
workflows.
Detailed hit-by-hit analysis. Every detected pattern with line numbers, severity tags, and explanations from the Antislop dataset.
slopsquid scan README.md
Quick density scores. One line per file showing score (0-100), rating, hit count, and word count for fast triage.
slopsquid score docs/*.md
Consolidated analysis for a website or directory. Per-page scores, aggregate stats, and the most frequent patterns across the corpus.
slopsquid report qry.zone
$ slopsquid scan test_doc.md
! test_doc.md — score: 68/100 (heavy)
[!!] line 3: "flickered" — 98.5% of 67 models
[! ] line 7: "murmured" — 73.1% of 67 models
[!!] line 15: "voice barely whisper" — trigram
18 hits in 42 lines, 387 words
$ slopsquid score *.md
! 68.0 heavy 18 hits 387 words test_doc.md
* 31.2 moderate 4 hits 892 words post.md
. 8.5 clean 1 hits 1204 words README.md
$ slopsquid report qry.zone
== SlopSquid Report ==
Source: https://qry.zone
Files scanned: 24 (3 skipped)
Breakdown: 18 clean, 4 moderate, 2 heavy
Average score: 14.2/100
Words like "flickered", "gaze", "murmured" that appear at abnormal rates across LLM outputs. Weighted by how many of 67 models overuse them.
Three-word sequences like "voice barely whisper" and "took deep breath" that are signatures of AI-generated prose.
Sentence-level constructions like "not just X, but Y" and hedging phrases that AI models produce at 2-6x human rates.
Every detection carries a weight from the Antislop paper. A word used by 98% of models scores higher than one used by 25%.
SlopSquid is a single Go binary with no external dependencies. Pattern data is embedded at compile time.
git clone
https://github.com/QRY91/slopsquid
cd slopsquid
go build -o slopsquid ./cmd/slopsquid
slopsquid scan README.md
slopsquid score docs/*.md
slopsquid report qry.zone