Scores your text from 0 to 5, flags every machine-written pattern, and can fail the build. Then a model from another family cleans the draft.

AI text has a telltale smell: delve, seamless, unlock, "it's not just a tool, it's a journey". Readers notice and stop trusting you. SlopMonster adds two things the other tools skip.
Five groups of patterns, one point each. Below 5/5, the command exits with an error, so it works as a CI gate. Developers call this a linter.
If Claude wrote it, GPT cleans it up. If GPT wrote it, Claude cleans it up. A model can’t hear its own accent. The script refuses to send the draft back to the same model family.
No “10,000 happy customers” when the business doesn’t have them. Any number next to a word for people gets flagged. If the number is real, --allow-proof becomes a warning.
The scorer gets the first and last word, because the scorer is honest and the model is persuasive. “Almost clean” is how a page ends up sounding just like every other AI page on the internet.
Kills delve, seamless, robust, unlock, elevate, leverage. Replace it with a simpler word, not a fancy synonym.
Kills not just X, but Y, reflexive lists of three, stacks of em dashes, piled-on caveats, and the closing summary nobody asked for.
One specific number per claim. Sentences of very different lengths. One edge case: a contraction, a fragment, a sentence starting with "And".
It matches one group and loses a point. Four groups strip away the AI accent. The fifth asks whether the line is selling anything.
Two lists. The first matches by root: elevate gets elevates e elevating. The second matches word for word, so it won't flag "we craft furniture by hand".
17 sentence patterns, not words. not just X, but Y, say goodbye to, that's where X comes in, and questions the author answers themselves.
Two em dashes within a 220-character window. Semicolons only above a threshold proportional to the page size.
faster, smarter, and better. Intentionally narrow: it only triggers on three loose words with an Oxford comma, or on a short sentence closing the clause.
A number next to a word for people, like 10,000+ happy users. Intentionally easy to trigger: a false alarm costs ten seconds; an error leaves a lie on your site.
The catalog is English-only. Portuguese text gets 5/5 because the scorer can’t read the language, not because it’s clean. Model-based cleanup works in any language.
The scorer is pure Python, uses the standard library, and has no dependencies. Cleanup uses the CLI codex or claude. Without either one, the script prints the prompt for you to paste into a chat.
Any recent version 3. The project CI runs on 3.11.
# check the version python3 --version
No installation. The scripts run directly from the folder.
git clone https://github.com/inematds/SlopMonster cd SlopMonster
For automatic cleanup: codex if you write with Claude, claude is written with GPT.
# one of the two, depending on the draft family which codex claude
All commands are real and run from the repo root. The step 1 example intentionally scores 3/5.
Pass the text directly. The scorer names each flag and exits with code 1 when it scores below 5.
python3 tools/deslop.py --text "It's not just a tool, it's a game-changing journey." # → score 3/5, flags both patterns, exit 1
In HTML, it reads only what visitors can see. In markdown, it skips code spans, fenced blocks, and strikethrough text, because literal text isn't sales copy.
python3 tools/deslop.py index.html python3 tools/deslop.py index.html --view hero # one element by id python3 tools/deslop.py README.md
The cleaned text goes to stdout, and notes about what changed go to stderr. Read the notes: sometimes the model deletes your closing line and only mentions it there.
tools/cleanse.sh rascunho.md > limpo.md 2> notas.txt # writing with GPT? make Claude the cleaner: DESLOP_WRITER=gpt tools/cleanse.sh rascunho.md > limpo.md
A cutting-edge model is great at removing markers and quite capable of adding new ones while it does so. Cleanup without re-scoring is a coin toss.
python3 tools/deslop.py --text "$(cat limpo.md)" # only publish at 5/5
If the number is true and you can prove it, rule 5 becomes a warning and still prints the matches for you to review.
python3 tools/deslop.py index.html --allow-proof
The repo workflow runs the tests and scores the published text on every push. Copy it to your repo and point it to the file with your text.
cp .github/workflows/slop.yml <seu-repo>/.github/workflows/ # in the yml, replace examples/gate-sample.txt with your file python3 tools/deslop.py --text "$(cat examples/gate-sample.txt)"
Copy the folder into skills and call it. The SKILL.md it's pure markdown, so it works for Codex and other agents too.
cp -r SlopMonster ~/.claude/skills/slopmonster/ # in Claude Code: /slopmonster, "remove the slop from this", "does this sound like AI?"
The catalog matches word roots, and a bad stemming shortcut leaves the scorer quietly half-blind. The test suite catches that.
python3 tools/test_deslop.py bash tools/test_cleanse.sh
Real lines from the Ridgeline Roofing site and an unedited pass on the Jasper.ai homepage, which went from 3/5 to 5/5 in one pass, with the meaning intact and nothing invented. Full records at examples/ no repo.



