The most useful thing in no-ai-slop is the naming. Everyone who reads a lot of model output can feel when a paragraph is machine-shaped, but very few people can say why in a way an editor can act on. This skill turns that feeling into a list of named patterns with an example of each, and then edits against them.
The README says 20 or more patterns are detected, and it prints eleven of them with an example each. Some are openings: throat-clearing ("Here's the thing..."), faux-insight setups ("What nobody tells you..."), colon reveals ("The best part: it learns."). Some are structures: binary contrasts ("It's not X. It's Y."), negative listing ("Not a X. Not a Y. A Z."), dramatic fragmentation ("That's it. That's the whole thing.").
The rest are the padding that fills a paragraph when there is nothing to say: superficial analysis ("...highlighting the team's commitment"), importance puffery ("marks a pivotal moment"), weasel attribution ("experts agree," "studies show"), fake-strong verbs ("serves as a centralized hub"), and synonym cycling, where the same thing becomes the agent, then the assistant, then the tool.
Read that table next to anything you have shipped recently. The synonym-cycling entry in particular is the kind of tell that can survive several rounds of self-editing, because each individual sentence looks fine.
Two modes, and the second one is the interesting one
Editing works the way you would guess: invoke the skill with a draft and you get the edited draft back plus a short section describing what changed. The stated policy is the minimum effective edit, which matters, because the failure mode of automated line editing is a rewrite that removes the author along with the slop.
The detection mode is the one I would reach for more often. Ask whether a piece reads as AI and you get back every pattern it found, each with the quoted line. The judgment stays with you, which is what you want when the writing is yours. It is also the mode that works on someone else's draft without the awkwardness of handing back a version you rewrote.
Beyond the pattern list, the skill enforces a handful of fundamentals: lead with the point when it helps, use active voice, untangle hard-to-follow sentences, and prefer concrete numbers over abstractions. My own editing bias is that the last one earns its place, because a sentence forced to carry a real number has nowhere to hide.
It checks its own edit
The workflow centers on two files. SKILL.md holds the editing rules and workflow, and eval.md holds pass and fail checks the skill runs on its own edits. So the loop is edit, then grade the edit against a fixed rubric in the repo, then report.
That is a small design decision with an outsized effect on trust. A style guide an agent reads once at the start of a long task tends to fade by the end of it; a check that runs on the finished output is harder to overlook. If you maintain house voice rules for model-written copy, the pattern is worth stealing whether or not you install this skill: keep the rules and the acceptance checks in separate files, and make the checks run last.
Installation is deliberately low-ceremony. You paste a line into Claude Code, Codex, or another harness asking it to install the skill globally from the repository, and the agent does the rest. License is MIT.
What it does not tell you
The README is short and does not pretend otherwise. There is no before-and-after example longer than the table entries, no measurement of how often detection fires on human writing, and no statement of which model it was tuned against, which matters because these patterns can shift from one model generation to the next. The author presents it as one skill extracted from a larger personal library, so its rules reflect one person's ear rather than a consensus style guide.
Even so, the pattern list is short enough to read in two minutes and specific enough to argue with, which is more than most style guidance manages. Install it or copy the table into your own review checklist; either way the naming is the part that transfers.
GitHub: https://github.com/petergyang/no-ai-slop
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)