Slopmonster

SlopMonster

Take any draft and make it read like a person wrote it. A landing page, a README, an email, a script. The target is not "passes a detector". Detectors are noise, and chasing them makes prose worse. The target is the gut of a reader who has seen a thousand AI paragraphs this month.

The loop is always the same four steps, and the linter gets the first and last word, because the linter is honest and the model is persuasive.

1. LINT      python3 tools/deslop.py --text "…"      score /5, exits red below 5
2. REWRITE   three passes, by hand or by model       (see below)
3. CLEANSE   a DIFFERENT model family strips tells   tools/cleanse.sh
4. RE-LINT   python3 tools/deslop.py again           ship only at 5/5

Step 1 — Lint

python3 tools/deslop.py page.html              # a built page (scores visible text only)
python3 tools/deslop.py page.html --view hero  # one element by id
python3 tools/deslop.py --text "paste a draft"
python3 tools/deslop.py page.html --allow-proof  # numbers are real and evidenced

Regex, no opinions. Five groups, one point each: AI vocabulary, AI constructions, punctuation cadence, rule-of-three rhythm, invented proof. Below 5/5 it exits non-zero, so it works as a build gate. "Mostly clean" is how a page ends up sounding like every other AI page on the internet.

Empty input fails rather than passing. A cleanse that times out leaves a zero-byte file, and a gate that stamps that CLEAN reports slop as clean exactly when the pipeline broke.

Touching a regex means running python3 tools/test_deslop.py. The catalogue is matched by word root, and the obvious stemming shortcut silently kills a dozen base words.

Step 2 — Rewrite (three passes)

Full catalogue in references/signs-of-ai-writing.md. The short version:

1. Kill the vocabulary. delve, seamless, robust, unlock, elevate, leverage, game-changing, journey, realm… Replace with a plainer word, not a synonym of the same word.
2. Kill the shapes. not just X, but Y is the single loudest tell in English right now. Also the rule-of-three reflex, em-dash pile-ups, hedge stacks, symmetrical paragraphs, the closing summary nobody asked for, and a bold lead on every bullet.
3. Put a person back in. Removing tells leaves clean, dead copy. One specific number per claim. Sentence lengths that vary hard. One thing a cautious writer would have cut. One rough edge — a contraction, a fragment, a sentence starting with "And".

Step 3 — Cleanse with a rival model

A model is bad at hearing its own accent. A rival model hears it instantly. So the cleanse runs on a different model family than the one that wrote the draft:

tools/cleanse.sh draft.md > cleansed.md                 # copy out, notes on stderr
tools/cleanse.sh draft.md > cleansed.md 2> notes.txt    # keep the notes as well

The instruction it carries (prompts/cleanse.txt): strip the tells, keep every fact, keep the length within 10%, invent nothing.

It returns two things. The rewritten copy, then a <<<SLOPMONSTER-NOTES>>> line, then up to five bullets naming each tell and its fix. The script splits them, so stdout is copy and stderr is notes, and the redirect above writes prose only.

Read the notes. PASS 2 tells the model to stop at the last real point, so it will sometimes delete your closing line, and the notes are the only place it says so.

WARNING no <<<SLOPMONSTER-NOTES>>> line means the model ignored the format and the whole reply came through as copy. Check the tail before you ship.

Step 4 — Re-lint

Always. A frontier model is very good at removing tells and quite capable of adding new ones while it does. An unlinted cleanse is a coin flip.

The one hard rule

Never invent proof. No user counts, no testimonials, no ratings, no trusted by 10,000 teams unless every one is true and you can show it. If a claim needs a number you do not have, write [needs number] and move on. The linter flags number-plus-noun patterns on purpose: a false positive costs ten seconds, a false negative is a claim you cannot back. Specificity beats borrowed credibility anyway.

Output format

Return the rewritten copy first, in full. Then a short ▎ what changed list — at most five lines, each naming the tell and the fix. Never return analysis alone. Keep the author's meaning exactly: de-slopping is not rewriting the argument. Match the register you were given.