PlaybookIntermediate13 min read13 September 2026

The Slop Cleanup Kit

Fix generic AI copy without erasing your voice.

In short

Start with one page. Save the original, collect the facts and examples of your real voice, run a quote-based audit, approve individual edits, then check the facts and page before publishing. Peter Yang’s No AI Slop is a free, MIT-licensed editing skill with detect and edit modes. Your chosen AI service may cost money. These checks improve copy; they do not prove authorship or guarantee sales, rankings or detector results.

Start with one page customers actually read

Choose your homepage, one important product page or one automated email. Generic copy often describes how impressive a business is without telling the reader what it sells, how it works or what to do next.

Your first deliverable is one improved draft with a record of every approved change. Keep the original. Do not replace live website copy directly from an AI response.

Copy the complete text into a document, including headings, buttons, footnotes and legal qualifications.
Write who reads it and the one action they should take.
Collect verified product facts, prices, delivery terms, guarantees and claim sources.
Add two short samples of your own writing that sound like you. Remove customer names, private messages and confidential details first.
Use an AI workspace approved for the information you share. Review its data-use settings before uploading business material.

Save three working files: Voice and Facts, Copy Audit, and Approved Copy. The templates below tell you what to put in each.

What counts as a useful warning sign?

Wikipedia’s Signs of AI writing is a descriptive field guide, not a universal test. Its editors explicitly warn that humans also use these patterns and that some signs are specific to Wikipedia. Neither a detector score nor an awkward sentence establishes who wrote a page.

For business copy, ask whether a pattern makes the page less specific, harder to understand or less trustworthy. A clear comparison, useful bullet list or distinctive phrase can stay. Fix the weak sentence, not every occurrence of a word.

The No AI Slop skill takes a useful approach: in detect mode it names the pattern, quotes the original line and gives a short fix without guessing authorship. In edit mode it asks for the minimum effective change and a summary of changes. Voice preservation is an instruction to the model, not a guarantee.

Twenty checks you can apply today

This is a business-copy checklist adapted from No AI Slop and Wikipedia’s field guide. It is not a scientific inventory of exactly twenty universal AI fingerprints. Examples below are invented to illustrate editing decisions, not claims about an actual business.

Check	What to flag	Smallest useful fix
1. Forced binary contrast	“It’s not a bag. It’s a lifestyle.”	Say what the bag is for. Keep a contrast only when it explains a real difference.
2. Empty opening	“In today’s fast-paced world…”	Delete the setup and begin with the customer’s task.
3. Fake insider knowledge	“What nobody tells you about shipping…”	State the shipping fact. Do not imply exclusive knowledge without evidence.
4. Dramatic colon reveal	“The best part: it works.”	Describe the mechanism or result directly.
5. Stacked fragments	“More clarity. More control. More growth.”	Combine into one useful sentence or replace with a verified benefit.
6. Repeated sentence shapes	Every paragraph starts “Whether you’re…”	Remove duplicate framing. Let the information set the rhythm.
7. Superficial analysis	“We added order tracking, highlighting our commitment to excellence.”	Keep “We added order tracking.” Explain its use only if needed.
8. Inflated importance	“A pivotal moment in the future of shopping.”	Name the actual change or delete the claim.
9. Unsupported authority	“Experts agree this is the best option.”	Name a relevant source and what it actually found, or remove the claim.
10. Decorative synonym switching	“The app… the platform… the solution…” for one product.	Use one clear name consistently.
11. Abstract business jargon	“Leverage a robust solution to optimize your workflow.”	State the task the product performs.
12. Empty verb phrases	“Has the ability to provide assistance with scheduling.”	Use “helps schedule” if that accurately describes the feature.
13. Lists of what it is not	“Not a tool. Not a service. A revolution.”	Say what the reader receives.
14. Self-answered rhetorical setup	“The result? Unmatched efficiency.”	Provide a supported result or describe the specific process.
15. Commentary about importance	“That last part matters more than you think.”	Explain the consequence, or let the preceding fact stand.
16. Generic profound ending	“The future isn’t coming. It’s already here.”	End on the next action or a concrete consequence.
17. Repeated conclusion	The last paragraph repeats every benefit.	Delete the recap and keep the useful next step.
18. Punctuation and formatting overload	Constant long dashes, decorative bold, emojis and headings over tiny sections.	Use punctuation for meaning and formatting for navigation. A dash alone proves nothing.
19. Generic praise replacing detail	“Innovative, premium quality for modern living.”	Restore the actual material, size, mechanism or policy from approved facts.
20. Invented evidence or unfinished output	Broken citations, nonexistent studies, “[insert testimonial]” or “as an AI…” in customer copy.	Open and verify each source. Remove placeholders and unsupported claims before publishing.

Priority: incorrect claims and misleading promises first; unclear information second; repetitive style third. Removing an awkward phrase does not repair a fabricated fact.

Use the skill without getting stuck on installation

The beginner route: paste the instructions

Open the repository linked at the very bottom of this guide.
Open the folder skills, then no-ai-slop, then SKILL.md. Open eval.md in the same folder too.
Copy both files into a fresh conversation in your approved AI assistant, or upload them as text files if the interface supports uploads.
Tell the assistant: “Use SKILL.md as the editing instructions and eval.md as the final review checklist for this conversation. Confirm you can read both before we begin.”
Paste your Voice and Facts template and the copy you want reviewed. Start with the audit prompt below.

The rest of this playbook is below.

Enter your email to read it. You also get one new playbook every week.

One email a week. Unsubscribe any time.

This loads the rules into the conversation. It does not permanently install a slash command. If the assistant cannot open GitHub or read attachments, paste the file contents yourself. A URL alone is not proof it read the instructions.

Install No AI Slop in your AI agent

No AI Slop is an editing skill, not a standalone app. Choose one installation method below. You need an agent that supports installing skills, such as a compatible Claude Code or Codex setup. Ordinary chat access does not automatically include permission to install files.

Before installing: Open the repository at the bottom of this guide and review its README, SKILL.md and eval.md. Global installation makes the skill available beyond one project for supported agents under your user account. On a managed work computer, get approval before adding third-party instructions or packages.

Option A: ask your agent to install it. Open your agent and paste the repository’s installation prompt below. This is an instruction for the agent, not a terminal command.

Install the /no-ai-slop skill globally from https://github.com/petergyang/no-ai-slop

Review any installation or file-access request before approving. Ask the agent to show the installed file path and confirm it can read the skill and evaluation rules. If it cannot install skills in that environment, use the paste-instructions method above instead. A message saying “installed” without a file or host confirmation is not enough.

Option B: install from your computer’s terminal. This route needs Node.js and npm, which provides npx. Open Terminal on macOS/Linux or PowerShell on Windows and check:

node --version
npm --version
npx --version

If a command is missing, install a supported Node.js LTS release from the official Node.js website or through your organization’s approved software manager, then reopen the terminal. Do not enter these commands into an ordinary chat window.

The repository documents this installation command:

npx skills add petergyang/no-ai-slop --skill no-ai-slop --global --yes

This downloads and runs the skills installer and installs the named skill globally for supported agents. The --yes flag accepts installer prompts automatically, so review the command and source before running it. Read the installation result to confirm which agent locations were updated. Do not assume every AI app on your computer now has the skill.

Verify before using business copy. Start a fresh agent session if the skill is not immediately visible. Ask it to load no-ai-slop and identify its installed location, then run this fictional test:

/no-ai-slop is this slop? In today's fast-paced world, our platform marks a pivotal moment in business innovation.

Check that the response quotes the supplied wording, names relevant patterns such as an empty opening or inflated importance, and suggests short fixes without rewriting the whole draft or assigning an AI-authorship score. Then ask it to edit the same sentence. It should not invent product features or statistics to replace the vague claim. If the command is unrecognized, check the target agent in the installer output or use the manual route.

Cost and privacy: The repository is free and MIT-licensed. Your AI service may charge for use. Copy submitted for review goes through your chosen AI host and its data policies; installing the skill does not make that processing local or private.

After a supported installation, the documented commands are:

/no-ai-slop is this slop? [paste your draft]

/no-ai-slop [paste your draft]

The first is detect mode; the second edits and lists changes. Replace the bracketed text with your actual copy. If your app treats the command as ordinary text, use the manual route above rather than repeatedly claiming it is installed.

The skill files are free under the repository’s MIT license. A paid assistant subscription, usage charges or organizational restrictions may still apply. Do not upload private copy to an unfamiliar third-party “humanizer.”

Create your Voice and Facts file

Copy this template into a document. Fill it before asking for an edit. If a field is unknown, write “unknown” and ask the model to flag it rather than fill it in.

ASSET: Homepage / product page / automated email
AUDIENCE:
READER QUESTION:
DESIRED ACTION:
VERIFIED FACTS: Features, materials, dimensions, prices, eligibility, delivery, returns.
CLAIM SOURCES: Fact ID | exact fact | source or approved policy | date checked.
NON-NEGOTIABLES: Legal text, warnings, exclusions, guarantees and exact numbers.
VOICE SAMPLES: Two short pieces I wrote and want to sound like.
VOICE TRAITS: For example, plain, mildly funny, direct, no fake urgency.
WORDS OR LINES TO KEEP:
THINGS WE DO NOT CLAIM:
ORIGINAL COPY:
[Paste the complete asset, including buttons and footnotes.]

Never invent a stronger number to make copy feel concrete. “Faster” cannot become “saves five hours” without evidence. Keep limitations that affect a purchase decision even when they interrupt a neat sentence.

Audit first, edit second

Prompt A: exact-line audit

Paste this after the instructions and your Voice and Facts file. Save the response as Copy Audit.

Use the No AI Slop instructions in detect mode. Do not rewrite the asset yet and do not assign an AI-authorship percentage or detector score.
Read the full original. First list the sections you received so I can check coverage.
For each real issue, return:
1. Location or heading.
2. Exact quote copied from my draft.
3. Named pattern or factual problem.
4. Why it hurts clarity, specificity or the reader’s decision.
5. Smallest suggested fix.
6. Facts or clarification needed before editing.
Prioritize unsupported claims, then missing or confusing information, then style.
Do not flag a line just because it contains a dash, a list or a common word.
Keep strong lines and briefly identify what should remain untouched.
Stop after the audit and wait for my approval.

Check before continuing: every quote must exist in the original. Reject invented quotes and subjective complaints that do not explain a reader problem. For long pages, work section by section and maintain a list of completed sections.

Prompt B: minimum-change edit

Mark each proposed change KEEP, EDIT, DELETE or NEEDS FACT. Then paste this prompt with those decisions.

Edit only the changes I marked EDIT or DELETE. Preserve my meaning, voice samples, structure and strong lines.
Use only facts in my approved fact file. Preserve numbers, prices, dates, units, guarantees, exclusions, product names and CTA destinations exactly unless I explicitly approve a change.
Do not invent examples, testimonials, urgency, sources or opinions.
If a fix needs a missing fact, flag it under Open questions and leave that claim unpublished rather than guessing.
Return:
A. The full revised draft.
B. A change log: original quote | replacement | reason.
C. Open questions and unresolved claims.
D. A brief check against eval.md, including any failed checks.
Do not publish or overwrite the original.

Prompt C: factual and voice review

Compare the original, approved fact file and revised draft.
List any changed or added number, policy, promise, qualification, feature, quote or source. Check each against the supplied evidence. Mark absent evidence as unverified.
Check whether my distinctive words, humor, uncertainty and useful details survived.
Flag new generic phrases, invented confidence and passages rewritten without a clear reason.
Confirm every original section was reviewed, including buttons, footnotes and email subject lines.
Return discrepancies and the exact passage to fix. Do not declare the draft accurate based only on your own earlier output.

Then do the source check yourself. Another AI pass can help find discrepancies, but it is not independent evidence. Read the revision aloud and compare it side by side with the original.

Three business-copy examples

All examples here are fictional. Each rewrite uses only the facts stated alongside it.

Homepage: replace praise with the actual job

Approved fact: The product puts enquiries from website forms and email in one inbox. No time-saving study exists.

Before: “In today’s fast-paced business landscape, our innovative platform empowers teams to unlock seamless communication.”

After: “See enquiries from your website forms and email in one inbox.”

Why: The rewrite names the inputs and output. It does not invent a revenue lift or time saving. A suitable button might be “See how the inbox works” if that destination exists.

Product page: retain personality and useful specs

Approved facts: A 750 ml bottle, dishwasher-safe body, lid washed by hand. Owner’s preferred line: “For people who forget where they left their water.”

Before: “For people who forget where they left their water. More than a bottle, it’s a testament to sustainable living. Its robust design elevates every sip.”

After: “For people who forget where they left their water. Holds 750 ml. Put the body in the dishwasher and wash the lid by hand.”

Why: Keep the writer’s joke. Replace generic claims with specifications and care instructions. Do not add an environmental claim without supporting evidence.

Automated email: get to the next action

Approved facts: The customer requested a demo. The booking link is valid. No deadline or scarce-slot claim is approved.

Before: “We’re thrilled to embark on this transformative journey with you. Take the next step toward unlocking your full potential.”

After: “Thanks for requesting a demo. Choose a time using the booking link below. Reply with the workflow you want us to show.”

Why: The email states why it arrived and what to do next. Keep required consent, unsubscribe and legal text unchanged.

Prevent the same problems in the next draft

Use this brief before generating new copy. Attach the completed Voice and Facts file.

Draft this asset using the attached Voice and Facts file.
Answer the reader’s main question early. Use specific approved facts and my voice samples. Keep meaningful limitations beside the claims they qualify.
Avoid generic openings, inflated importance, unsupported authority, forced contrasts and repeated dramatic fragments.
Do not replace missing evidence with confident language. Put missing information in a separate questions list.
Keep the CTA specific to the real next step. Do not invent scarcity or testimonials.
After drafting, audit for named patterns and make only changes that improve clarity. Return the draft and a short list of factual checks still needed.

Before approving a page, try the portability test: could the sentence appear unchanged on a competitor’s site? If yes, replace it with a supported detail or delete it. Do not delete useful navigation or standard legal wording merely because it is common.

Keep three reusable outputs

File	Contents	Done when
Voice and Facts	Audience, action, real writing samples, verified facts and prohibited claims.	Another editor can tell what must remain true and what the business sounds like.
Copy Audit	Asset location, exact quote, issue, proposed fix, approval and unresolved evidence.	Every change points to the original and an explicit decision.
Approved Copy	Final text, change log, source checks, approver, original backup and publication date.	Claims and links are checked, placeholders removed, and a human has signed off.

Review new or substantially changed copy before publishing. Recheck prices and policies whenever the underlying business terms change. Preserve the old version so you can restore it.

Troubleshooting

Problem	What to do
The rewrite sounds more generic.	Restore the original, provide stronger voice samples and restrict edits to the approved audit rows.
The model flags every sentence.	Require a specific reader problem for each flag. Delete vague findings. No rewrite is required when the original works.
It adds a statistic or stronger promise.	Reject the change. Supply a source or retain the narrower supported statement.
It changes a return policy or warning.	Restore the approved language and have the responsible owner review any proposed legal change.
Only the start of a long page was reviewed.	Split the page at headings, track section coverage, then do a whole-page consistency review.
The slash command does nothing.	Confirm host support and installation. Otherwise paste the skill and evaluation files into the conversation.
A detector still flags the edited page.	Check facts, specificity and readability. Do not add errors, fake anecdotes or random wording to manipulate the score.
The text is still unconvincing after cleanup.	Check whether the offer lacks proof, clear terms or a useful benefit. Editing cannot supply a missing product advantage.

Final publishing check

The reader can identify what is offered, who it is for and what to do next.
Every number, testimonial and consequential claim has a checked source.
Prices, guarantees, exclusions, care instructions and legal text are preserved.
The writer’s useful voice remains. No fake personal experience was added.
Buttons and links lead to the right places. The copy works on mobile.
No placeholders, model commentary or broken citations remain.
A human approved the text and saved the original before publication.

Track the business outcome separately from the style audit. If you test two versions, define one relevant outcome such as completed enquiries or purchases and keep other changes controlled where possible. A before-and-after change alone does not prove the copy caused a conversion difference.

What this changes

Your team can approve edits by checking the quoted sentence, the supporting facts and the proposed change. Keep that record with each published asset so the next round starts from approved copy instead of another generic rewrite.

Further reading: Wikipedia: Signs of AI writing. This guide adapts its observations for business copy; it is not Wikipedia policy or an authorship detector.

Original skill: No AI Slop by Peter Yang. Review the current README, skills/no-ai-slop/SKILL.md, eval.md and MIT license before use. The practical templates and fictional business examples in this guide are adaptations, not official examples from the repository.

Get the repository: https://github.com/petergyang/no-ai-slop