Open source · local scoring
The anti-slop editor for AI drafts.
Zero Slop scores the AI accent, rewrites the draft in your voice, and checks every fact against the original.
npx skills add manavmishra/ZeroSlop --global
Public benchmark: 50 synthetic drafts, a blind LLM-as-a-judge review, and two pooled rounds. Competitive here does not mean a general winner. Source snapshot 3790a1f
A rewrite can strip the stock phrases and still fail: faked candor, chopped rhythm, an invented detail. Zero Slop fixes the wording and leaves what the draft says alone.
Switch formats to compare an AI-heavy draft with its rewrite. The claim survives. The padding does not.
We’re thrilled to announce that our team has leveraged cutting-edge AI to deliver a seamless onboarding experience, cutting setup time by 40%. This milestone underscores our unwavering commitment to innovation and operational excellence.
We cut onboarding setup time by 40% using AI.
In today’s rapidly evolving landscape, AI is enabling content teams to accelerate the drafting process while maintaining authenticity. In this article, we’ll explore how thoughtful review workflows can bridge the gap and empower writers to create meaningful content.
AI helps content teams draft faster without losing their voice. This article looks at review workflows that protect authenticity and help writers produce work that matters.
To unlock scalable value, the organization should leverage a comprehensive, cross-functional transformation framework that aligns AI-enabled operating models with strategic priorities.
This integrated approach will empower teams to optimize decision-making and drive sustainable impact across the enterprise.
The organization should use one cross-functional plan to align its AI operating models with its priorities.
That plan should help teams make better decisions and create lasting value.
AI is transforming how we write, but here’s what nobody tells you: the future isn’t about replacing writers.
It’s about leveraging AI to accelerate the drafting process while unlocking the unique human creativity that makes great writing possible. A thread 🧵
AI can speed up drafting without replacing writers.
Human creativity still makes great writing possible.
The sample scores come from the open-source Zero Slop scorer. Each rewrite keeps the main claim from its draft.
How it works.
Local Python measures traceable evidence. Your host model edits the draft in context. Scripts and the model verify the exact text returned.
Local Python
Finds traceable surface evidence.
A 0 to 100 score. Every point traceable.
Host language model
Interprets the draft before changing it.
Claude, GPT, or another compatible model.
Scripts plus host model
Checks the exact text you receive.
Any change sends the text through every check again.
Separate diagnostics: a host-model predictability probe, and a portfolio probe for repetition across 3+ related drafts. Neither changes the surface score.
A cut span, a kept flag, a preferred fix. Human feedback becomes evidence.
Recurrence, novelty, and known-human safety. One edit cannot activate a rule.
Detector weights and fix preferences. Later evidence reconfirms; stale guidance decays.
This does not retrain the host model or perform reinforcement learning or RLHF. It updates an inspectable private detector and rewrite memory.
Every point traces back to a phrase or document statistic. The result is a heuristic editing signal, not an authorship probability.
Pattern meter
The scorer checks 267 weighted patterns, a 96-word watchlist, and 25 context-gated terms. Those terms count only when the surrounding sentence makes them useful evidence.
Rhythm
Followability
The scorer flags comma pileups, clusters of long words, and sentences of 38 words or more.
Formatting and register
Emoji, hashtag clusters, heavy bolding, repeated dashes, and machine-formal language count when they cluster.
A separate host-model predictability probe can sit beside the score. It is reported on its own and never folded into the surface meter.
Human corrections become feedback, but one edit pair cannot change the meter.
A cut may be a missed tell. A flagged phrase you keep may be a false positive.
Phrases need 3 content-distinct pairs. Single words need 5, plus novelty and human-corpus safety.
The private overlay adapts both detection and fixing. Later evidence reconfirms it; stale guidance fades.
The benchmark covers 50 AI-heavy drafts in 6 kinds of writing, compares four tools, and uses a blind LLM-as-a-judge review. The repository includes both the data and the study's limitations, so you can check the headline yourself.
Read the benchmark notesCross-genre regression.
18/18The set contains three paraphrases for each of six modules: LinkedIn, X, email, blog, newsletter, and research. They were informed by anonymous public examples, not copied from them. This is regression coverage, not real-world accuracy.
This pinned audit compares documented capabilities, not writing quality. It shows where Zero Slop adds measurement, verification, cross-draft diagnosis, and private learning around the rewrite.
| Capability | Zero Slop | blader/humanizer | petergyang/no-ai-slop |
|---|---|---|---|
| Detect-only evidence report | Native | Not documented | Guided |
| Traceable numeric surface meter | Native | Not documented | Not documented |
| Statistical rhythm and format signals | Native | Guided | Guided |
| Scripted fact and claim check | Native | Guided | Guided |
| Best-of-N candidate reranking | Native | Not documented | Not documented |
| Separate copy desk and read-aloud gates | Native | Guided | Guided |
| Cross-draft portfolio probe | Native | Not documented | Not documented |
| Private evidence-gated learning | Native | Not documented | Not documented |
| Recurring rewrite-fix memory | Native | Not documented | Not documented |
| Evidence reconfirmation and decay | Native | Not documented | Not documented |
| External voice sample or profile | Native | Guided | Not documented |
| Public outcome regression harness | Native | Not documented | Not documented |
Repository audit of the pinned commits. Native means a dedicated script, stored artifact, or named workflow gate. Guided means an instruction or self-check without a dedicated executable component. Not documented means the capability was not found in the audited repository; it does not prove the product cannot do it.
Capability presence is not effectiveness proof. The blind review found Zero Slop and blader/humanizer competitive in this setup (p = 0.15).
Open the audit dataA plausible-looking draft can save the sender a few minutes while handing the reader a harder job: reconstruct the intent, check the facts, and repair the result. The evidence below comes from different settings, so each figure keeps its method and limits attached.
Survey signal
Workers who received so-called workslop estimated nearly two hours to resolve each incident. Forty percent said they had received it in the prior month. BetterUp and Stanford study
Controlled study
In METR's early-2025 experiment, experienced developers predicted a speedup but took 19% longer with AI. It was one narrow setting, not a verdict on every developer or newer tool. METR study
Adoption signal
Gallup's May 2026 indicator shows a gap between tools being integrated and employees using them regularly. Clear plans are rarer still. Gallup indicator
Among the 31,476 developers who answered that survey question, 45% also said debugging AI-generated code takes more time. The setting is software, but the failure mode is familiar: fluent output still needs expert verification. Stack Overflow survey
In the BetterUp and Stanford survey, the cleanup shifted effort across the team and changed how recipients viewed the colleague who sent it. Survey source
RAND cites external estimates of more than 80%. Its own interviews traced failure to problem definition, workflow fit, data, infrastructure, and technology-first planning. Sloppy output can add review debt, but it is not the whole explanation. RAND report
One em dash does not make a draft machine-written. Formal prose is not automatically AI-made. Zero Slop weighs patterns, not isolated signals.
Before a rewrite passes, Zero Slop checks every figure, name, quote, and link for additions or omissions.
Add a sample of your writing so the scorer can recognize which habits belong to you.
Star the project, report something that feels wrong, or send Zero Slop to the person who always ends up editing the team's AI drafts.
Zero Slop runs in Codex, Claude Code, Cursor, Gemini CLI, OpenCode, Warp, and Zed, plus any agent that reads SKILL.md files.
npx skills add manavmishra/ZeroSlop --globalZero Slop's scoring, rewrite, and copy-desk instructions
Compatible agents
What the tool does, where it runs, and what it costs.
AI slop is writing that falls into familiar model habits: safe phrasing, even rhythm, template structure, stock transitions, and polished sentences that say very little. Zero Slop measures those surface patterns and shows you exactly where they occur.
No. Detectors estimate whether a machine wrote something. Zero Slop gives the draft a traceable heuristic surface score, then edits the writing and checks figures, names, quotes, and links against the original. The score is not an authorship probability.
Zero Slop is designed to prevent that. It inventories figures, names, quotes, and links, then rejects a rewrite that drops one or adds a new one. You still make the final editorial call.
No. The scorer is a local Python script that uses only the standard library. It works offline and needs no account, server, or network connection. The skill itself runs inside the coding or writing agent you already use.
Zero Slop works with agents that support SKILL.md files, including Codex, Claude Code, Cursor, Gemini CLI, OpenCode, Warp, and Zed. The repository also documents setup for ChatGPT and claude.ai.
Yes. Zero Slop is open source under the MIT license. The scorer uses only Python's standard library; it needs no third-party packages.
Zero Slop can compare the draft it returned with the version you publish. One edit pair cannot change the meter. A phrase needs the same signal across 3 content-distinct pairs; a single word needs 5. Eligible evidence must also be new and safe against the human-writing reference set. Approved changes stay in a private local overlay and affect the next run.
The website records aggregate page traffic, performance, chapter views, scroll depth, example choices, and clicks on install, GitHub, feedback, and sharing controls. No cookies, persistent identifiers, fingerprinting, or draft text are collected. Global Privacy Control and Do Not Track disable the custom interaction events.
Lower is better.