A free, MIT-licensed skill file called stop-slop teaches Claude to detect and strip the robotic phrases and structural patterns that mark AI-generated prose, doing in seconds what a human editor or Grammarly Business subscription does for $25 to $150 per use.
stop-slop is a free, MIT-licensed Claude skill file that scans any piece of AI-generated prose and removes the tell-tale phrases, sentence rhythms, and structural habits that mark it as machine-written, doing the same work that Grammarly Business charges around $25 per seat per month for, or that a human AI editor bills at $50 to $150 an hour to do manually.
The problem it solves is one every marketing team knows but rarely talks about directly. You have a content team that generates blog posts, case studies, LinkedIn updates, and sales emails with AI. The drafts come back fast. But they also come back with a specific texture: opening sentences that throat-clear before saying anything, sentences structured as binary contrasts ("It's not about X, it's about Y"), a fondness for adverbs like "truly" and "ultimately," and a narrator-from-a-distance voice that treats the reader as a subject being observed rather than a person being addressed. These patterns erode trust in ways readers can feel even when they cannot name them.
The standard remedies are either expensive or slow. Grammarly Business has grown into a platform that catches this class of problem through tone and clarity scoring, but it costs real money at team scale and does not specifically target the AI-tell problem. Human editors who specialize in humanizing AI content are charging a premium precisely because demand is outrunning supply. And asking the original LLM to "make this sound less AI-generated" tends to produce a slightly different version of the same problems.
stop-slop is a different approach. It is not software you install. It is a structured instruction set, written as a Claude skill, that you load into your AI workflow once and then apply to any draft as a review pass. The skill contains a curated list of banned phrases (everything from "At its core" to "It's worth noting that"), a catalog of structural cliches to catch (the rhetorical setup followed by a pivot, the dramatic sentence fragment used for emphasis, the passive voice that hides who did what), and a scoring rubric across five dimensions: directness, rhythm, trust, authenticity, and density. Any draft that scores below 35 out of 50 goes back for revision with specific line-level notes.
Setup is genuinely simple. If you use Claude Code, you add the repository as a skill. If you work in Claude Projects, you upload the SKILL.md file and its reference files to the project knowledge base. If you are calling the API directly, you include the skill in your system prompt. There is no database to configure, no server to run, and no subscription to manage.
That said, honesty about the limitations matters here. stop-slop is not standalone software. It requires an LLM, which means it requires an API key and incurs token costs every time you run a review pass. Those costs are small at normal content volumes, maybe a few cents per article, but they are not zero. The project has twelve commits and a single maintainer, which means there is no support contract, no SLA, and no guarantee that the phrase list will stay current as AI writing patterns evolve. The skill is also tuned for English-language prose and has not been validated for other languages or technical writing contexts where some of the flagged patterns might be appropriate.
The business case is clearest for teams that are generating AI content at volume and paying either for Grammarly Business seats, for freelance editing time, or both. A mid-sized marketing team running twenty to thirty AI-generated pieces a month through a human editor for humanization review is spending real budget on a task that stop-slop can handle as an automated first pass, shifting the human editor's time toward judgment calls that actually require judgment, the structural arguments, the factual accuracy, the audience fit, rather than phrase-level cleanup.
It is also useful as a forcing function for teams that are not yet thinking systematically about AI content quality. The phrase list alone is worth reading. Seeing "leveraging" and "seamlessly" and "at the end of the day" laid out together in a banned list makes it obvious why AI-generated content can feel exhausting to read: it uses every shortcut at once. The skill makes those patterns visible and actionable instead of leaving it to a reader's vague sense that something is off.
The repository has accumulated 7.5k stars in a short window, which suggests the problem it addresses is landing with people building AI content workflows, not just researchers or hobbyists. The author, Hardik Pandya, published the skill publicly and licensed it MIT, meaning any team can fork it, extend the phrase list, and adapt the scoring rubric to their own house style without asking permission.
The question stop-slop cannot answer for you is whether your readers care enough about AI tells to justify the friction of adding a review pass. For some audiences and some content types, they do not. For audiences where voice and authenticity carry real weight, the question answers itself.
There is something quietly instructive about the fact that the solution to AI writing feeling like AI writing is, itself, more AI, configured more carefully.