Most human readers sense when text was produced by a model even when grammar and facts are correct — predictable openers, emphasis crutches, and rhythmic clichés give it away. Stop Slop treats those artifacts as a targetable style problem: instead of rewriting for meaning, it trains an LLM to hunt and remove the specific patterns that make prose sound machine-generated.
What Sets It Apart
- Focused scope: rather than fixing grammar or clarity, the repo enumerates “AI tells” (phrases.md), structural clichés (structures.md) and paired examples (examples.md) so an LLM can perform surgical edits that preserve content while removing detectable markers. This makes it complementary to grammar/style tools.
- Skill-first packaging: core guidance lives in SKILL.md and is intended to be included in system prompts or uploaded as a Claude skill, enabling direct integration into assistant workflows without needing new orchestration layers.
- Measurable output: a 5‑dimension scoring rubric (Directness, Rhythm, Trust, Authenticity, Density) and per-dimension 1–10 ratings give editors a quantitative threshold to decide when to revise further.
- Practical rule set: sentence-level constraints (avoid Wh- starters, ban certain punctuation/fragmentation, prefer active voice) plus concrete banned-phrase lists make the transformations predictable and auditable.
Who It's For and Trade-offs
Great fit if you run content pipelines that use LLMs but need text to read like a human writer (content ops, in-house communications, prompt engineers). It speeds bulk sanitization and provides a repeatable audit trail via examples and scores. Look elsewhere if your priority is creativity, rhetorical flourish, or preserving idiomatic voice in non-English languages: aggressive pattern removal can flatten stylistic nuance and may over-correct culturally specific phrasing. Also, it is a style-sanitizer, not a substitute for domain editing or fact-checking.
Where It Fits
Think of Stop Slop as an intermediate layer between a generator and final human review: it reduces obvious model fingerprints so editors can focus on content quality and correctness. It pairs well with existing style guides and grammar checkers but is distinct because it targets generative artifacts rather than mechanical errors.
