These four open-source skills run inside an AI assistant and give it different editing instructions. Your choice depends on whether you want help removing generic phrasing, preserving source details, or reviewing an edit. Results also depend on the assistant you use.
In a replay of our AI Slop test corpus, Zero Slop had the lowest mean score on its own meter. That is a narrow finding, not an independent endorsement. We build Zero Slop and publish the test. Use the results to choose what to trial, then review the actual edits.
Compare the editing workflows
Start with the kind of help you need. These are documented capabilities, not a new performance test. Each tool still relies on your assistant for the contextual edit.
| Editor & source | Editing approach | Checks described |
|---|---|---|
| Zero Slop | Assistant edits, then reviews grammar, spoken flow, and the final text. | Local Python scoring and tracked source-detail checks, alongside the assistant's contextual review. |
| avoid-ai-writing | Separate rewrite, detect-only, and targeted file-edit modes. | Optional local detector and preservation validator for protected content and structure. |
| no-ai-slop | An editing checklist with separate edit and detect modes. | Detect mode asks the assistant to quote patterns without scoring the draft or identifying its author. |
| humanizer | For pasted text, a rewrite, critique, and final version by default. | The assistant reviews remaining patterns. File and embedded modes return only the final text. |
The links point to the exact instructions we reviewed. Check each edit against the original text for changes in meaning, and review cost and privacy before sending confidential text.
Four instruction sets on the same drafts
The starting mean was 76.3. A lower score means fewer patterns on Zero Slop's meter; it is not a probability of AI authorship. The release checks also assess paragraph structure, so the pass count reflects more than the mean score.
| Editor | Mean score ↓ | Rewrites passing release checks | Rewrites keeping tracked details | Word count change |
|---|---|---|---|---|
| Zero Slop | 12.8 | 100% | 100% | -8.9% |
| avoid-ai-writing | 23.3 | 83.3% | 100% | -14.6% |
| no-ai-slop | 28.4 | 66.7% | 94.4% | -13.7% |
| humanizer | 35.4 | 50% | 94.4% | -7.2% |
Run recorded August 26, 2026, using the historical instructions pinned in the saved records. Check current releases separately. Inspect the pinned run records and raw outputs or read the benchmark overview.
Try Zero Slop with a draft you know well, then compare the edit against your original. Prefer working in your own assistant? Install the skill.
Which editor fits your workflow?
The editors are listed in the result order above. The workflow recommendations below are editorial judgments based on documentation reviewed September 17, 2026.
Zero Slop
mean 12.8 · 100% of rewrites passed the release checks
Edits checked against the original
Consider Zero Slop when you need to compare an edit with its source. Its local tools check tracked names, numbers, quotes, and links; your AI assistant interprets context and rewrites the prose.
Watch for: A rewrite can keep every number and still change who did what. The automatic checks cannot establish that meaning survived, so read the final edit against the source.
Visit Zero Slop's repository ↗avoid-ai-writing
mean 23.3 · 83.3% of rewrites passed the release checks
A structured audit with detect and edit modes
The project's documented workflow separates rewriting, detection, and targeted file edits. It also provides an optional local detector and preservation validator. Consider it when you want an audit of the patterns before accepting a rewrite.
Watch for: It shortened the drafts most in this replay. That may suit a tight brief, but fewer words are not automatically a better edit. Check whether explanations your reader needs were removed.
Visit avoid-ai-writing's repository ↗no-ai-slop
mean 28.4 · 66.7% of rewrites passed the release checks
A readable checklist you can inspect and adapt
Peter Yang's skill documents patterns such as inflated significance, staged openings, and recycled contrasts. Its detect mode quotes passages without trying to identify their author. Consider it if you want to understand the editing instructions yourself.
Watch for: A checklist still needs context: a rhetorical contrast or a repeated opening can be intentional. In this replay, one draft failed the tracked-detail check; inspect the output before generalizing.
Visit no-ai-slop's repository ↗humanizer
mean 35.4 · 50% of rewrites passed the release checks
A visible critique when you paste a draft
For pasted text, Humanizer's default is a rewrite followed by a critique and a final edit. It distinguishes personal writing from neutral reference prose and instructs the agent not to invent missing facts. Consider it when seeing the revision process helps you decide what to keep.
Watch for: Its outputs had the smallest net reduction in word count in this replay, but the highest mean score on our meter. The current project may differ from those pinned instructions. This result is not a test of its latest release.
Visit humanizer's repository ↗
First Reader or Zero Slop?
Choose a tool for the review you need.
- First Reader
- Explore where a draft might lose its audience. Two simulated readers react passage by passage; a separate skimmer checks the opening. Notes and attention strips sit beside the draft. It does not rewrite.
- Zero Slop
- Find generic phrasing, revise it, and compare the result with the source. Local checks track names, numbers, quotes, and links. Read-aloud and fresh-eyes passes review the complete edit.
First Reader is not included in the rewrite replay of our AI Slop test corpus: editing is outside its documented scope. Simulated reactions are hypotheses, not measured human behavior. Source-detail checks are not proof that meaning survived.
Source review: September 10, 2026 (Pacific). First Reader at f56f4feb. Its helpers run locally; model cost and privacy depend on the assistant you use.
What the replay can and cannot establish
Our AI Slop test corpus contains deliberately obvious drafts across six genres, chosen to check for recurring faults. Each workflow used GPT-5.4 at high reasoning, in batches of three, with pinned instructions. The saved records identify the instructions, drafts, prompts, and outputs by file hash. You can inspect the run, but asking the model again may produce different text.
The same model and drafts reduce some differences between runs. They do not remove our selection bias, the scorer's preferences, or variation in model output. Our deliberately difficult examples target recurring faults. Results apply to this selected corpus; the replay does not estimate quality across everything people write. There is no independent human-preference ranking here.
Read the columns together. A tool can lower a phrasing score by deleting useful context. A detail check can confirm that a name or number survived without establishing that the surrounding claim is still true. In this replay, avoid-ai-writing removed more words than Zero Slop while keeping the tracked details in every draft. Whether that is preferable depends on what those words were doing.
We have not measured editing time saved, uptime, throughput, or a statistically reliable population-wide quality advantage in this comparison. The replay also does not test AI-detector evasion. Inspect an edit even when its score is low; publish only after reviewing it.
Repeat the comparison on your own work
Write down what would make an edit unusable before running a trial. For a research summary, that might be a missing limitation. For instructions, it might be a changed prerequisite or command. Choose drafts you have permission to share with the assistant, and keep an untouched copy of each source.
| Draft to include | What to check in the edit |
|---|---|
| A draft you would already publish | Does the tool leave clear wording alone? Record unnecessary changes and any loss of your voice. |
| A result with a caveat | Do the population, measure, time period, and uncertainty survive? The fictional pilot example provides a source and an incorrect rewrite to inspect. |
| Instructions with an exception | Check the order of actions, prerequisites, commands, links, and who must do each step. Reject an edit that changes the procedure. |
- Choose drafts before seeing the outputs. Include one you already like, so unnecessary rewriting is visible.
- Keep the source, model, editing brief, and model settings fixed. Save each tool’s instructions with its revision, output, and run date.
- Score the saved outputs with the same scorer revision. Then check for changed claims, omitted caveats, and shifts in voice, without using the number as the verdict.
- Record failures as well as improvements. If possible, ask a reader who does not know which tool produced each edit to compare them.
Give each workflow the same editing task, followed by its own documented invocation. A reusable brief is: improve this draft for its stated audience, preserve the claims and qualifications, leave clear passages alone, and flag anything that needs evidence. Save the brief with the outputs so you can distinguish a tool difference from a prompt change.
For each output, record whether you would accept it, repair it, or discard it, with the source passage that explains the decision. If time matters to your choice, measure the review and repair work yourself. This comparison has not measured that cost. A tool that needs fewer corrections on your actual documents may fit better than the tool with the lowest score here.
Choose a workflow only after its output meets your factual and privacy requirements. Keep the decision tied to the tested model and revisions; rerun the trial when either changes. You can share the saved method and examples without claiming that a small personal trial establishes a general product ranking.
The replay directory includes the method, evaluator, output files, and run records. Our source-checking guide explains what to look for when a fluent sentence changes the claim.
Cost and privacy depend on your assistant
All four projects publish their skill source. That lets you inspect the instructions; it does not make the AI assistant running them free, private, or offline. Check the project's license and your assistant's plan, retention terms, and organization policy before sending confidential text.
Zero Slop's local scorer runs on your device without sending the draft. Its hosted editor sends text for AI rewriting and is capacity-limited. A locally installed skill can still send text through its host assistant. Those are different data paths; choose the one your document permits. See the privacy policy and the sandbox's local mode.
Commercial alternatives, not benchmarked
QuillBot offers free and paid paraphrasing options. Grammarly offers free and paid writing assistance. Consider their documented workflows if you want a packaged writing product. Neither hosted product was run in this same-model replay, so neither receives a score or rank here.
Free and paid availability was checked September 17, 2026. Follow the provider links for current quotas, billing terms, and supported features.
Before you choose
Does an AI humanizer detect authorship?
No. A rewriter changes text; an authorship detector estimates who or what produced it. Zero Slop checks writing patterns and source details, not authorship. Neither a rewrite nor a low slop score guarantees a detector result.
Can I use the results to choose a winner?
Use them to shortlist workflows for your own drafts. The table is a historical, same-model regression result measured with our checks. It does not establish a universal winner across genres, languages, models, or current product versions.
Which option should I try first?
For source checks, try Zero Slop. For a structured detect-and-edit workflow, inspect avoid-ai-writing. For a readable pattern checklist, inspect no-ai-slop. For a rewrite with an explicit critique, inspect Humanizer. Compare the actual edits on one draft you know well.
Will a cleaner score make the writing better?
Not necessarily. A useful technical term, deliberate repetition, or a precise qualification may trigger a pattern. Keep it when it serves the reader. Choose a usable edit over a smaller number.
Need a workflow rather than a product ranking? See five anti-slop workflows, from a local first check to a final source review.