Removing Emphasis Crutches from AI-Generated Content with Stop-Slop

Stop-Slop is a data-only editorial skill that strips emphasis crutches and AI-writing artifacts by injecting curated markdown rulebooks directly into an LLM’s system prompt or post-processing pipeline.

Removing emphasis crutches from AI-generated content requires more than simple keyword filtering—it demands structured editorial intelligence. The hardikpandya/stop-slop repository provides a portable, code-free "skill" that eliminates predictable LLM prose habits—including throat-clearing openers, business jargon, and filler constructions—by supplying eight core editorial rules and comprehensive reference dictionaries that any model can consume.

Core Architecture and File Structure

Stop-Slop is organized as a self-contained knowledge package containing zero executable code. Every component is a markdown file describing editorial policy, making it compatible with any LLM platform that supports system-prompt injection or file-based knowledge imports.

The Master Rule Set (SKILL.md)

The central instruction set resides in SKILL.md, which enumerates eight high-level editorial rules designed to detect and remove emphasis crutches. The file contains:

• Rules 1-8 covering active voice enforcement, filler phrase elimination, and formulaic structure breaking
• A quick-check checklist for real-time evaluation (e.g., "Any adverbs? Kill them.")
• A scoring rubric where outputs scoring below 35/50 require revision

Reference Libraries

Three specialized dictionaries in the references/ directory provide the lexical and structural inventory for identifying AI slop:

• references/phrases.md: A concrete list of emphasis crutches and filler phrases to delete, including entries like "Here's the thing:", "Full stop.", and "lean into"
• references/structures.md: Pattern catalogue of prose constructions to avoid, such as binary contrasts ("Not because X. Because Y."), dramatic fragmentation, and false agency
• references/examples.md: Before-and-after transformations illustrating how rules apply to real text

How the Skill Works in an LLM Pipeline

The cleaning process follows a five-step evaluation cycle:

1. Load Core Rules: Inject SKILL.md into the system prompt or knowledge base
2. Reference Consultation: Direct the LLM to check phrases.md and structures.md when triggered
3. Quick-Check Execution: Evaluate each sentence against the checklist during generation
4. Compliance Scoring: Run the draft against the scoring table; revise if below 35/50
5. Iterative Refinement: Continue rewriting until the threshold is met

Because the repository supplies both low-level lexical items (specific crutch words) and high-level structural patterns (rhetorical constructions), the LLM performs surgical edits without over-pruning meaning.

Three Ways to Implement Stop-Slop

System Prompt Injection

For API-driven workflows, fetch SKILL.md at runtime and prepend it to your system prompt:

import requests

# Load the skill dynamically

skill_url = "https://raw.githubusercontent.com/hardikpandya/stop-slop/main/SKILL.md"
skill_content = requests.get(skill_url).text

system_prompt = f"""
You are Claude. Load the Stop-Slop skill:

---BEGIN SKILL---
{skill_content}
---END SKILL---

When you generate text, obey all rules in the skill.
"""

# Send to your LLM client

response = claude_api.generate(
    system_prompt=system_prompt,
    user_prompt="Explain the benefits of micro-frontends."
)

Claude Projects Knowledge Base

For persistent access across multiple conversations:

1. Clone the repository locally
2. Upload SKILL.md and the entire references/ directory to your Claude Project’s knowledge section
3. Reference the skill in your system prompt: Load skill from project: stop-slop

Claude automatically reads the files and applies the rules during generation, consulting the emphasis crutch lists in phrases.md when drafting responses.

Offline Post-Processing

For batch cleaning without API costs, apply the lexical rules directly using Python:

import re
import pathlib

def remove_emphasis_crutches(text: str) -> str:
    """Strip known crutches using the phrase inventory."""
    phrases = pathlib.Path("references/phrases.md").read_text()
    
    for line in phrases.splitlines():
        if line.startswith("- "):
            phrase = re.escape(line[2:].strip('"'))
            text = re.sub(r'\b' + phrase + r'\b', '', text, flags=re.I)
    
    # Clean up residual whitespace

    text = re.sub(r'\s{2,}', ' ', text)
    return text.strip()

raw = """Here's the thing: building products is hard. Not because the technology is complex. Because people are complex. Let that sink in."""
clean = remove_emphasis_crutches(raw)

print(clean)

# Output: "building products is hard. technology is complex. people are complex."

This method leverages the references/phrases.md inventory to remove emphasis crutches without invoking an LLM, ideal for high-volume content pipelines.

Summary

• Stop-Slop is a data-only editorial skill residing in the hardikpandya/stop-slop repository, distributed under the MIT license
• The system relies on SKILL.md for eight core rules and a 50-point scoring rubric where scores below 35 trigger revision
• Emphasis crutches are catalogued in references/phrases.md, while structural patterns live in references/structures.md
• Implementation requires no code execution—only markdown injection into system prompts or knowledge bases
• Offline processing is possible by regex-matching against the phrase inventories for API-free content cleaning

Frequently Asked Questions

What exactly are emphasis crutches in AI writing?

Emphasis crutches are predictable rhetorical tics that LLMs use to simulate conviction or transition, such as "Here's the thing:", "Full stop.", or "Let that sink in." According to the Stop-Slop source code, these phrases function as filler that signals AI authorship without adding semantic value, and they are catalogued for removal in references/phrases.md.

How does the Stop-Slop scoring rubric work?

The rubric in SKILL.md assigns points based on compliance with eight editorial rules, including active voice usage and absence of throat-clearing openers. A score of 35 out of 50 is the minimum threshold; anything below requires iterative revision until the text meets the cleanliness standard.

Can I use Stop-Slop with models other than Claude?

Yes. Because the repository is data-only markdown with no executable dependencies, you can inject SKILL.md into any LLM supporting system prompts—including GPT-4, Llama, or Gemini—or use the phrase lists for offline regex processing in Python.

What is the difference between phrases.md and structures.md?

phrases.md contains lexical emphasis crutches—specific words and short phrases to delete—while structures.md catalogs higher-order prose patterns like binary contrasts ("Not because X. Because Y.") and dramatic fragmentation that require rewriting rather than simple deletion.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}