no-ai-slop Strips 20+ AI Writing Tells From Any Draft. That Doesn't Make It Yours.

Peter Yang's open-source skill catches the em-dashes and the "It's not X. It's Y." Here's what it's good for, and why a clean scan isn't the same as good writing.

no-ai-slop reads a draft and hands you back the same text with the machine scrubbed out. It ships a table of the tells it hunts, and the table is the interesting part: binary contrasts like "It's not X. It's Y.", throat-clearing openers like "Here's the thing," colon reveals like "The best part: it learns," weasel attribution like "experts agree," and dramatic fragmentation like "That's it. That's the whole thing." More than twenty patterns, each with the smell it leaves behind.

The skill hit number two on the trending board this week, and the ranking isn't the story. The story is what has to be true for a de-slopping tool to trend at all. Enough model-written prose is shipping right now that a dedicated pass to remove the smell of it is a utility people star and install. The demand is the signal, not the novelty.

A de-slopper is a detector you can write against. It is not a ghostwriter.

Here's my position, and the whole article hangs off it. Stripping the fingerprints of a machine does not add the fingerprints of a person. Run a hollow draft through no-ai-slop and you get a cleaner hollow draft. The em-dashes are gone, the stock verbs are gone, the fake-profound ending is gone, and the thing still has no author, because the tool can tell you what to remove and it cannot tell you what to say.

That's not a knock on the skill. It's the correct way to use it. As a detector, it's sharp and honest, and it makes you a better editor of your own work. As a substitute for having a point, it produces text that passes every check and means nothing. The difference between those two uses is the entire game.

What it actually is

no-ai-slop is MIT-licensed, written by Peter Yang, who runs the Behind the Craft newsletter, and as of this week it sits at roughly 1.6k stars and 122 forks (repo). It's small on purpose. Two files do the work: SKILL.md holds the editing rules and workflow, and eval.md holds pass/fail checks the skill runs against its own edits before it hands anything back. You install it by pasting one line into Claude Code, Codex, or another agent harness: "Install this skill globally," followed by the repo URL.

It runs in two modes, and both matter. In edit mode you paste a draft after /no-ai-slop and get back a revised version plus a short "What changed" section, so you can see every cut rather than trusting a black box. In detect mode you ask "is this AI slop?" and get every pattern it found, each with the quoted line it found it on. Detect mode is the one I'd reach for first, because it's a mirror rather than a rewrite.

The pattern list is smarter than the usual banned-words file, and the reason is that it names smells instead of tokens. "Importance puffery" catches "marks a pivotal moment." "Fake-strong verbs" catches "serves as a centralized hub." "Synonym cycling" catches the paragraph that calls the same thing the agent, then the assistant, then the tool, to avoid repeating a word. "Negative listing" catches "Not a database. Not a cache. A memory." These aren't words you can grep for. They're shapes, and shape-level detection is what makes the skill worth more than a find-and-replace.

The good part is the four rules nobody talks about

Under the detector, no-ai-slop enforces four fundamentals that make writing good: lead with the point when it helps, use active voice, untangle hard-to-follow sentences, and prefer concrete numbers over abstractions. That last quarter of the skill is the part I'd tape to the wall. It has nothing to do with AI. It's just what good editing was before models existed, and it's most of the value in the whole repo.

The engineering choice worth stealing is eval.md. The skill checks its own edit against explicit pass/fail criteria before returning it, which means it has somewhere to fail instead of grading its own vibe. That's the same discipline the better agent skills have started shipping: don't let the model vouch for itself, give it a rubric with teeth. A de-slopping skill that de-slops its own output and then verifies the result against a checklist is doing more real work than most one-shot editing prompts.

Where the model can't reach

The tells no-ai-slop catches are surface tells. Removing a filler verb doesn't hand you a point of view, and killing a colon reveal doesn't give you something worth revealing. The skill operates on the top layer of the writing, and the top layer is not where slop actually comes from. Slop comes from having nothing to say and asking a model to fill the space anyway. You can scrub that draft spotless and it stays empty.

Yang seems to know this, and his own workflow is the tell. He writes a first draft by hand, iterates with a model on edits, then does another manual pass at the end. The tool sits in the middle of a human sandwich. Used the other direction, generate a draft, de-slop it, ship it, you get text that clears every pattern and has no author behind it. The skill is the same in both cases. The result is not.

Put this into practice

You can get most of the value in the next ten minutes without changing how you write.

Run detect mode on something you already published. Paste an old post and ask /no-ai-slop is this AI slop? Read the quoted lines it flags. Some will sting because they're right and some will be false alarms on choices you made on purpose. Both are useful. This is the mirror, and it costs nothing.
Steal the four fundamentals and skip the install if you want. Lead with the point, active voice, untangle the sentence, concrete numbers over abstractions. That's the bulk of the benefit and it needs no tool. Write those four on a sticky note and edit against them by hand.
If you edit with it, keep the sandwich. Your draft first, the tool in the middle, your own pass last. The order is the whole safeguard. Never run it the other way, where the model writes and the skill sands off the evidence.
Watch for over-correction. The README says it makes the "minimum effective edit," which is the right instinct, but any de-slopper will occasionally flag a deliberate fragment or an honest colon as a defect. When it wants to smooth out a line you wrote on purpose, you're allowed to say no. You're the author; it's the copy editor.

Where it bites

The limits are worth naming plainly. It catches surface tells, not empty thinking, so it can make bad writing look professional without making it good. It will sometimes flag real style, because a fragment or a colon can be a choice rather than a slip. It's one skill pulled from a larger personal system that points at a paid newsletter and courses, and while the skill itself is genuinely MIT and free, the funnel is part of the design. It's shaped for English prose and won't map cleanly onto other languages or onto code comments.

There's a bigger risk that shows up only at scale. If everyone de-slops against the same twenty patterns, the patterns become the new house style. You strip out the machine's tells and, along the way, the small irregularities that made a given writer sound like themselves. A million clean drafts, all clean the same way. The tool that removes sameness can manufacture a subtler sameness if the whole internet runs it.

The uncomfortable question it leaves

no-ai-slop is a good detector and a bad ghostwriter, and that gap is the entire point of it. It can find the machine in your writing. It cannot put a person there. Only you can do that, and the tool just assumes you already have.

So run it on your own last post before you decide how you feel about it. Read the lines it quotes back at you. Then ask the real question the whole trend raises: if a utility for removing the smell of AI writing can climb to number two in a week, how much of what you read this week was written by someone who had nothing to say and a model that was happy to say it for them?

Sources: no-ai-slop repo and README, SKILL.md, Trendshift board, Peter Yang's write-up on the skill.