Catch the tells that make writing read like an AI wrote it, and rewrite them to a genuine human voice. A free checker built on Claude Code that reviews your prose, names each cliché, hedge, and formatting quirk with a concrete fix, and cleans up drafts as they are written. It is a craft tool for better writing, not a way to beat a detector.
/plugin marketplace add sigistry/marketplace/plugin install slop-check@sigistry/slop-check:audit ./README.mdPowered by the slop-check plugin.
Ten categories of tell, from the cliché words that cluster in AI prose down to the leftover "as an AI" a chat window pastes in. Findings are ordered by what matters: structure and register first, then phrasing, then punctuation.
delve, leverage, tapestry, robust, seamless, pivotal, underscore, realm. Flagged by density in a passage, not a single use, so good writing that happens to use one is left alone.
"it is important to note", "in today’s fast-paced world", "plays a crucial role", "a rich tapestry of". Filler that adds words, not meaning.
"not just X, but Y", "it’s not X, it’s Y", "no X, no Y, just Z". The shape LLMs reach for to sound insightful while correcting a belief the reader never held.
"serves as", "stands as", "functions as", "acts as" standing in for a plain "is" or "are". The plainer verb is almost always better.
Three adjectives or clauses where one carries the load, repeated across a piece. Clean, elegant, and intuitive becomes clean.
Travel-brochure adjectives and unearned significance: vibrant, nestled, groundbreaking, a testament to, underscores its importance. Say what happened instead.
"could potentially", "aims to", "experts argue", "studies show" with nothing cited. Either state what is, or name the source.
Em dashes standing in for other punctuation, decorative emoji, bold on every other phrase, Title Case On Ordinary Headings, divider lines between every section.
Sentence after sentence of the same length and shape. Human writing varies it. A short sentence after two long ones resets the ear.
"As an AI", "as of my knowledge cutoff", "I hope this helps", and stray citation fragments like oaicite or [cite: 1] pasted from a chat window.
Good writers use em dashes, emoji, and the word delve. So a linter that flags every one of them is wrong more often than it is right, and it flattens strong writing into gray mush. slop-check weighs concentration and register instead: several markers stacked together, in a flat and promotional voice, in a rigid shape. That is what it flags, and it leaves a well-placed dash alone. It also ignores AI-detector scores completely, because chasing them makes prose worse, not better.
Add the marketplace and the slop-check plugin to Claude Code.
Claude applies the rules as it writes, or you run /slop-check:audit on any file to review it after the fact.
You get each tell with its location, the reason it reads as AI, and a concrete rewrite. Nothing is changed on disk unless you ask.
| AI Slop Checker | AI "humanizer" tools | AI detectors | |
|---|---|---|---|
| Reads your actual draft | Yes | Rewrites it on a server | Only scores it |
| Names the cause of each tell | Yes | No | No |
| Suggests a concrete rewrite | Yes | Auto-rewrites, often vaguer | No |
| Works while you write | Yes | No | No |
| Goal is clearer writing | Yes | A lower detector score | A detector score |
| Free, no signup | Yes | Usually paid | Usually paid |
It is a free tool that runs inside Claude Code. You install the slop-check plugin, and it reviews or drafts prose so it reads like a person wrote it rather than an AI. Run /slop-check:audit on a file and it lists each tell with a concrete rewrite; while Claude is writing, it applies the same principles so the first draft comes out clean.
AI slop is the set of habits that make text read as machine-generated: overused words like delve and leverage, stock phrases like "in today’s fast-paced world", a rigid intro-three-points-conclusion shape, a flat promotional register, uniform sentence rhythm, and pile-ups of em dashes and emoji. No single one is proof; the tell is a cluster of them together.
No, and that is deliberate. slop-check is a craft tool for clearer writing, not a way to fool a detector. Detector scores are noisy and gaming them produces worse prose, so the plugin ignores them entirely. If you want your writing to read well, this helps; if you want to evade detection, this is the wrong tool.
Humanizer tools rewrite your text on their servers to lower a detector score, often making it vaguer in the process. slop-check does the opposite: it reads your actual draft, names each specific issue with the reason it reads as AI, and suggests a sharper rewrite you can accept or ignore. It runs locally in Claude Code, and it is about voice and clarity, not detection.
No. That would be its own kind of slop. Used well, an em dash or an emoji is fine, so the default review judges by clusters and register and leaves a well-placed one alone. It never changes your meaning or drops citations. For contexts that want zero tolerance, a strict house style is available (no spaced em dashes, no decorative emoji, ASCII quotes).
Yes. The slop-check plugin is free and open source (MIT). There is no signup, subscription, or hosted dashboard. It runs locally in Claude Code on your own machine.
The highest-signal ones: cliché words in clusters (delve, leverage, tapestry, robust), the "not just X, but Y" construction, copula avoidance ("serves as" for "is"), promotional adjectives (vibrant, nestled, groundbreaking), stock phrases ("it is important to note"), uniform sentence length, and overuse of em dashes and emoji. slop-check checks for all of these and explains each call.
Both. Because the skill’s description covers writing and editing, Claude loads it on its own when it recognizes it is drafting a README, a post, or release notes, and it applies the principles as it goes. You can also run /slop-check:audit on any existing file to review it after the fact.