stop-slop
Installation
SKILL.md
Stop Slop
Slop is text that has the shape of writing without a writer making choices. Every sentence is grammatical and nothing is wrong, yet a reader feels within two lines that nobody meant it. This skill removes that feeling by fixing what causes it, not by deleting a list of words.
The core idea: patterns are symptoms
Each LLM tic has a cause. Delete the surface pattern and the cause grows a sibling: ban "delve" and you get "dig into"; ban em dashes and you get a colon in every sentence; ban "it's not X, it's Y" and you get "X? No. Y." So diagnose the cause, then fix the cause.
| # | Cause | What it produces | The fix |
|---|---|---|---|
| 1 | Pull toward the most probable phrasing. A model predicts likely continuations, so it drifts to the phrase that fits the most contexts. Specifics are improbable; generalities are safe. | Generic claims, clichés, abstract nouns ("enhance the overall experience"), stock metaphors (journey, landscape, tapestry), brochure adjectives. | Specificity. Replace the category with the instance: the number, the name, the actual thing that happened. |
| 2 | Rewarded for looking helpful at a glance. Preference training uses raters who compare answers quickly. Structure, completeness, and enthusiasm win at a glance. | Headers, bold, bullets, triads, recaps, "comprehensive" coverage, exclamation-level enthusiasm. | Write for a reader who reads every word. Prose by default; structure only when the content is a list or a reference. |
| 3 | Penalized for being wrong or one-sided. Safety and accuracy training make every claim cheaper to qualify than to commit to. | Hedge stacks ("may potentially"), reflexive both-sidesing, "it's important to note", disclaimers. | Commit. Hedge only where there's real uncertainty, once, and say what is uncertain. |
| 4 | No stake and no taste. The model doesn't know which detail matters, so it marks everything as mattering, and it has no position of its own. | Significance inflation ("pivotal", "a testament to", "transformative"), narrator-from-nowhere voice, no opinions. | Decide what matters and say why. Let small things be small. Cut the rest. |
| 5 | Absorbed genre templates. Training data is heavy with marketing copy, SEO posts, LinkedIn, school essays, TED-style talks. | "In today's fast-paced world", "Here's the thing:", "The result?", "Let's dive in", "In conclusion", hook-then-reveal. | Name the template and step out of it. Start with the point. |
| 6 | Sentence-by-sentence generation with a weak model of the reader. Each sentence tries to close itself; the model can't tell the reader already got it. | Paragraphs ending in a moral, "-ing" tails that restate significance, signposting, recaps, explaining the obvious. | Trust the reader. End the paragraph when its content ends. |
| 7 | Trained-in repetition avoidance and uniformity. | Synonym cycling ("the company... the firm... the tech giant"), relentless parallelism, metronome rhythm, or its newer opposite: staccato fragments for drama. | Repeat the key noun. Let rhythm follow the content. |
| 8 | The assistant persona. The text was produced as a reply to someone. | "Great question!", "I hope this helps", "Let me know if...", "Certainly!", flattering warmth. | Remove the chat frame. The text is not a reply. |