SlopMonster scores writing for AI tells and asks a rival model to fix it, 401 stars in 30 days

ItsssssJack's SlopMonster is a linter for AI-written prose. It scores a piece of writing on five categories of AI tells, uses a red exit code below a perfect score of five out of five, and then optionally asks a rival model to rewrite the flagged parts before scoring again. The repository is at 401 stars, released under MIT, and the scorer runs on the Python standard library with no dependencies.

The five detectors are named in the README. AI vocabulary catches words like delve, leverage, seamless and unlock. AI constructions catches shapes like "not just X, but Y" and "that's where X comes in". Punctuation cadence catches multiple em dashes inside one sentence. Rule-of-three rhythm catches three adjectives in a row, such as "faster, smarter, and better". Sales and marketing catches invented proof, such as fake user counts or claims a reader cannot verify. The scorer reads plain text, HTML (visible content only) and Markdown (it skips code spans, fenced blocks and struck-through text).

The rival-model rule

If the linter fails a piece, the cleanse script asks a different model family to rewrite the flagged parts and then re-scores. The README states the rule: if Claude wrote the draft, the cleanse routes to GPT-5.6; if OpenAI wrote it, the cleanse routes to Claude. The script refuses to send a draft back to its own family. The stated reason is that a model rewriting its own copy tends to produce the same tells with different words.

What the numbers describe

The one worked example in the README is a run on seven sentences from Jasper.ai's live homepage. The original scored 3 out of 5. After one cleanse loop it scored 5 out of 5, kept the length within 10 percent of the original, and did not add facts the source did not contain. That is one page, from the author's own testing, not an independent benchmark; the README does not report a larger sample.

The scorer has no dependencies, so it drops into a CI job that has Python and no more. The cleanse script needs one AI CLI on the runner, or none if the pipeline only wants the score. The catalogue is English-only.

The reason a team building software would look at this is that writing copy for a marketing page, a release note or a support reply is now often the same pipeline as writing code: a model drafts, a person edits, something ships. A pre-commit or pre-publish check that fails the build on the same AI tells a careful reader would notice makes that pipeline behave like the code one. Whether the specific catalogue matches a given house voice is a separate question, and both the detector rules and the vocabulary list are plain text files a reader can edit.

Source

ItsssssJack/SlopMonster on GitHub. Star count and windowed traction from the candidate feed on 2026-09-18. Detector categories, scoring scale and cleanse rule from the repository README.