You can spot AI prose in a paragraph. It announces itself before it says anything, leans on the same phrases, and ends with a punchy one liner that sounds profound and says nothing. Hardik Pandya's open source repo stop-slop attacks that pattern directly: it is a skill file that teaches an LLM to catch and remove recognizable AI tells, with a short rulebook, a banned phrase list, and a 50 point scoring rubric. As of October 26, 2026, it has 18,351 stars and 1,353 forks on GitHub, and it is MIT licensed.
Key Takeaways
>
Stop-slop is a single skill file plus three reference lists: banned phrases, structural patterns to avoid, and before/after examples.
Its 8 core rules ban throat-clearing openers, emphasis crutches, business jargon, adverbs, passive voice, and vague declaratives, and require active voice, specifics, and varied rhythm.
Its scoring rubric rates a draft 1 to 10 on directness, rhythm, trust, authenticity, and density. Below 35 out of 50, you revise.
You can use it as a Claude Code skill, a Claude Projects upload, copied custom instructions, or a system prompt for API calls.
It is a rulebook, not a detector: it edits prose by rules, it does not measure whether a human wrote the original text.
The problem: AI prose has a recognizable shape
AI writing fails in predictable ways, and readers learn the pattern fast. A draft opens with "Here is the thing" or "It turns out", then announces that something "matters because", stacks business jargon like "lean into" and "deep dive", and closes with a sentence fragment for drama. None of those moves is wrong on its own. Together, they form a voice that sounds like a model imitating confidence rather than a person making a point.
The word for this low quality output even entered the dictionary: Merriam-Webster chose "slop" as its 2025 Word of the Year, defining it as digital content of low quality produced in quantity by means of artificial intelligence (Merriam-Webster). Stop-slop treats that voice as a checklist problem. The skill file, SKILL.md, does not ask the model to "write better" in general. It gives the model specific things to delete, specific structures to break, and a quick check list to run before delivering prose. That specificity is the whole idea: a vague instruction produces vague prose, while a banned list produces edits you can verify.
Watch the 30-second repo signal: Kill AI writing cliches with stop slop.
What the skill actually contains
The repo is small on purpose. The README describes a folder with one core file and three references:
SKILL.md: the core instructions, 8 rules, a quick check list, and the scoring rubric.references/phrases.md: phrases to remove, grouped by category.references/structures.md: structural patterns to avoid.references/examples.md: before/after transformations that show the rules in action.
The phrase list covers 8 categories: throat-clearing openers, emphasis crutches, business jargon, adverbs, meta-commentary, performative emphasis, telling instead of showing, and vague declaratives. The jargon table is practical rather than preachy: it pairs each word to avoid with a plain replacement, such as "navigate" with "handle", "unpack" with "explain", and "circle back" with "return to". The adverb rule is the bluntest in the file: kill all adverbs, no "-ly" words, no softeners, no intensifiers, no hedges.
The structural list, summarized in the README, targets binary contrasts, negative listings, dramatic fragmentation, rhetorical setups, false agency, narrator-from-a-distance voice, and passive voice. The sentence level rules add three more hard constraints: no sentence starters with a Wh- word, no em dashes, and no staccato fragmentation. Every sentence needs a human subject doing something, and inanimate objects may not perform human actions.
The 8 core rules, in plain language
The SKILL.md core rules read like an editor's red pen, not a style guide:
Cut filler phrases: throat-clearing openers, emphasis crutches, and all adverbs.
Break formulaic structures: binary contrasts, negative listings, dramatic fragmentation, rhetorical setups, and false agency.
Use active voice: every sentence needs a human subject doing something, with no passive constructions.
Be specific: name the specific thing, and stop letting lazy extremes like "every", "always", and "never" do vague work.
Put the reader in the room: "you" beats "people", and specifics beat abstractions.
Vary rhythm: mix sentence lengths, prefer two items over three, end paragraphs differently, and use no em dashes.
Trust readers: state facts directly, and skip softening, justification, and hand-holding.
Cut quotables: if a sentence sounds like a pull-quote, rewrite it.
The quick check list turns those rules into a pre-delivery pass: any adverbs, any passive voice, any inanimate thing doing a human verb, any Wh- starter, any "here is what" opener, any "not X, it is Y" contrast, three consecutive sentences at the same length, a punchy one liner ending a paragraph, an em dash, a vague declarative, a narrator-from-a-distance line, or a meta-joiner. Each hit gets an edit, not a debate.
The 50 point score
Stop-slop scores a draft on five dimensions, each rated 1 to 10:
Dimension | Question to ask |
|---|---|
Directness | Statements or announcements? |
Rhythm | Varied or metronomic? |
Trust | Respects reader intelligence? |
Authenticity | Sounds human? |
Density | Anything cuttable? |
The total runs to 50 points, and the rule is simple: below 35 out of 50, revise. That threshold matters because it stops the skill from being a pure deletion machine. A draft can lose every banned phrase and still score low on rhythm or density, and the rubric forces a second pass on the shape of the prose, not just its vocabulary.
Before and after: what the edits look like
The examples file shows the rules working on short passages, and the transformations are more useful than the rules alone. One example takes "Here is the thing: building products is hard. Not because the technology is complex. Because people are complex. Let that sink in." and turns it into "Building products is hard. Technology is manageable. People are not." The opener, the binary contrast structure, and the emphasis crutch all go, and the direct statements stay.
Another example cuts a jargon stack down hard. "In today's fast-paced landscape, we need to lean into discomfort and navigate uncertainty with clarity. This matters because your competition is not waiting." becomes "Move faster. Your competition is." The point survives, the scaffolding does not. A third example removes hedging and permission granting: "It turns out that most teams struggle with alignment. The uncomfortable truth is that nobody wants to admit they are confused. And that is okay." becomes "Teams struggle with alignment. Nobody admits confusion."
Those examples also show the skill's limits honestly. They are short passages, edited for punch. Long reports, technical documentation, and legal text need hedges, qualifiers, and precise adverbs sometimes, and a rule that says "kill all adverbs" will overreach there. Stop-slop works best as a first pass on marketing copy, blog drafts, and general prose, with a human deciding which cuts to keep.
How to use it
The README gives four ways in, depending on your setup:
Claude Code: add the folder as a skill.
Claude Projects: upload
SKILL.mdand the reference files to project knowledge.Custom instructions: copy the core rules from
SKILL.mdinto your own instructions.API calls: include
SKILL.mdin your system prompt, with reference files loading on demand.
That flexibility is part of why the repo spread. It is not tied to one product's skill format alone, even though it uses the Agent Skills file shape: a SKILL.md file with YAML front matter carrying a name and a short description, plus reference files beside it. Agent Skills became an open standard in October 2026, when Anthropic donated the name, logo, GitHub organization, and agentskills.io domain to the Agentic AI Foundation under the Linux Foundation, with more than 70 platforms integrating the format. A skill like stop-slop rides that format: one file, portable rules, no installer.
If you already use agent skills across tools, the same idea pairs naturally with a shared skill setup: Agent Scripts: Share One Set of AI Coding Skills Across Claude Code and Codex.
The honest caveats
Stop-slop is a rulebook with sharp edges, and using it blindly will make prose worse in some places. "Kill all adverbs" deletes words that sometimes carry real meaning, such as "slowly" in a safety instruction or "directly" in a technical claim. Banning every Wh- sentence starter removes legitimate questions. Banning em dashes entirely is a style choice, not a grammar rule, and plenty of strong human prose uses them. The skill's own scoring step is manual: a model rates its own draft 1 to 10 per dimension, which is a judgment call, not a measurement.
It also does not detect AI text. If you need to know whether a passage was machine written, this repo will not tell you. It only rewrites a draft so it carries fewer of the tells that readers associate with AI output. Used with a human editor in the loop, that is a real improvement. Used as an automatic "humanizer" to disguise machine output, it is just a better mask, and readers will still notice when the ideas underneath are thin.
Why this repo matters
The repo's popularity says something about where AI writing stands right now. At 18,351 stars, 1,353 forks, and only 12 commits since its creation on January 11, 2026, stop-slop is not a fast-moving codebase. It is a short document that names a frustration many writers already felt but had not listed. Readers do not dislike AI prose because it is machine made. They dislike it because it announces, hedges, and performs instead of saying the thing. A checklist that deletes the announcement, the hedge, and the performance leaves the thing itself, and that is usually enough.
If you write with AI tools, try the skill on your next draft and score the result on the five dimensions before you publish. Keep the cuts that make the point clearer, restore the qualifiers that carry real meaning, and watch which banned phrases keep sneaking back into your first drafts. Those repeat offenders are your real tells.
The repo is worth reading in full, even if you never install it: hardikpandya/stop-slop.
Read this article in Indonesian: Stop-Slop: Skill Open Source yang Menghapus Cliche AI dari Tulisan Anda.
FAQ
What is stop-slop?
It is Hardik Pandya's open source skill for removing AI tells from prose, published at github.com/hardikpandya/stop-slop. It contains a core skill file, a banned phrase list, a structural pattern list, and before/after examples.
How popular is the repo?
As of October 26, 2026, it has 18,351 stars, 1,353 forks, and 12 commits on GitHub. It was created on January 11, 2026, and is MIT licensed.
How do I use stop-slop?
Add the folder as a Claude Code skill, upload SKILL.md and the reference files to a Claude Project, copy the core rules into your custom instructions, or include SKILL.md in your system prompt for API calls.
What does the scoring rubric measure?
Five dimensions, each rated 1 to 10: directness, rhythm, trust, authenticity, and density. The total runs to 50 points, and a draft below 35 out of 50 should be revised.
Does stop-slop detect AI generated text?
No. It rewrites prose by rules so it carries fewer AI tells. It does not measure whether a human or a model wrote the original text.
Is banning all adverbs and em dashes good advice for every text?
No. Those rules work well as a first pass on general prose, blog drafts, and marketing copy. Technical, legal, and safety writing often needs qualifiers, precise adverbs, and punctuation that the rules would delete, so a human editor should restore what carries real meaning.