AI Quality

Stop your writing from sounding like AI.

Zero Slop scores your writing 0–100, points at the exact words dragging it down, rewrites them, and double-checks that every fact stayed in and nothing was invented.

Install in one line

npx skills add manavmishra/ZeroSlop --global

Then say “de-slop this” in any agent. No account, no server, your writing never leaves your machine.

Why it matters now

Sounding like AI now costs you reach.

LinkedIn, Snapchat, YouTube and Substack all started acting on slop in mid-2026. A flagged post stops reaching anyone outside your own network. When up to 40% of social writing is machine-made, the systems ranking it bury anything that reads the same way.

Who it is for

Anyone who writes under their own name.

A founder writing a launch post

A marketer shipping five things a week

A researcher who needs their paper to stay formal

An engineering team that wants slop to fail a check the way a bug does

How it works

Scores, rewrites, and verifies.

01

Measure

Scores your draft 0–100. Every point traces to a phrase you can see — never a black box. Word choice, sentence rhythm, readability, formatting weight.

02

Rewrite

Writes several versions of flagged passages and keeps the cleanest one. No model of its own — borrows whichever one you are already in.

03

Verify

Lists every figure, name, quote and link in your original. Fails a rewrite that loses one or adds one.

Architecture

The full pipeline.

Every stage is traceable. Every score points back to a phrase you can see.

MEASURE · every channel · 3 of 4 blind to wording
wording
Pattern meter
74 tells · 55 lexicon · 13 riders
blind
Rhythm
burstiness · uniformity · shape
blind
Followability + format
density · dashes · register
Predictability
the model channel · uses the host model
Evidence fusion
0–100, every point traceable
Diagnose
hollow spans · facts · voice
Two-pass rewrite ×N
strip, build, then rerank
VERIFY · all three must pass
Numeric gate
likelihood, burstiness
Readalong
reads aloud, flags stumbles
Fidelity
no fact added or lost
fail → retry ≤ 3×   |   pass → text · scorecard · heatmap
REFLECT · what you changed teaches the meter
Your edit
cut a span → candidate tell
kept a flagged one → false positive
Three gates
3 documents agree · not already caught · clears the human corpus
learned.json
patterns up, weights down, stale tells decay out
TUNE · SkillOpt rewrites the instructions to score higher
SKILL.md
the rewrite instructions
Run + reward
run drafts · reward.py: fidelity + de-slop
best_skill.md
kept only if the val score rises
commit the winner · gated on a held-out split

Results

Head to head, it wins.

50 AI-heavy drafts, six kinds of writing, judged blind. Across 100 picks, judges chose Zero Slop 55 times to the next best at 40.

55/100
blind judge picks
~1,100
documents scored per second
0
facts lost in rewrites
74
tests passing

What makes it different

Not just another humanizer.

A score you can trust and check

Built on rules you can read, not a black box. Every point traces to a phrase, and the score can go straight into a CI build.

A promise about your facts

Most humanizer tools invent details or drop numbers to make sentences flow. Zero Slop lists every fact in your original and fails a rewrite that changes one.

An agent that learns you

After it hands your draft back and you edit it, it compares the two. A phrase only becomes a new rule after three separate people cut it.

Self-improving instructions

The rewrite follows instructions tuned by SkillOpt. It keeps an edit only when it beats the old version on drafts it has never seen.

Open source

MIT licensed. 74 tests. View the code, run it yourself, contribute.

Stop sounding like a machine.

One line to install. One phrase to run. Your writing stays on your machine.

npx skills add manavmishra/ZeroSlop --global