How to Use Stop Slop Reference Files (phrases.md and structures.md)

Stop Slop reference files are markdown lookup tables loaded via SKILL.md that tell language models which filler phrases and structural patterns to prune from text, applying prescribed fixes until the prose is clean.

The hardikpandya/stop-slop repository ships a declarative "skill" for cleaning AI-generated writing. At its core, two markdown files—phrases.md and structures.md—provide human-readable, version-controlled guidance that any compatible environment can consume to enforce direct, human-like prose.

What Are Stop Slop Reference Files?

Stop Slop uses static markdown files as configuration rather than code. These files live in the references/ directory and define the exact patterns the model must target.

phrases.md

references/phrases.md contains a curated list of filler phrases, weak adverbs, and meta-commentary that should be removed or rewritten. Each entry pairs a prohibited construction with a concise remediation instruction.

structures.md

references/structures.md catalogs recurring sentence-level rhetorical patterns—such as binary contrasts ("Not because X, because Y") and negative listings—that force the model to rewrite indirect setups into direct statements.

How the Reference Files Work

Both files are referenced from the main skill definition SKILL.md at the repository root. The first core rule in SKILL.md points to references/phrases.md, while the second core rule points to references/structures.md.

When loaded, the engine applies these rules as lookup-style filters:

1. Identify a match – The model scans input text for lines matching entries in phrases.md (e.g., "Here's the thing:") or structural patterns described in structures.md.
2. Apply the prescribed fix – Each entry includes an "Instead" recommendation. For binary contrasts, the instruction is "State Y directly." For filler adverbs, it is "Kill the adverb."
3. Iterate – The model re-runs the check until no further entries apply, guaranteeing a comprehensive clean-up pass.

Loading Reference Files in Your Projects

Because the files are plain markdown, you can ingest them into any workflow that consumes text: Python prompt builders, Claude skills, or shell scripts.

Python Implementation

Fetch the raw files and embed them directly into a system prompt:

import requests

BASE = "https://raw.githubusercontent.com/hardikpandya/stop-slop/main"

def load_md(path: str) -> str:
    resp = requests.get(f"{BASE}/{path}")
    resp.raise_for_status()
    return resp.text

phrases_md = load_md("references/phrases.md")
structures_md = load_md("references/structures.md")

system_prompt = f"""You are a text-cleaner. Follow the rules defined below.

--- PHRASES TO REMOVE ---
{phrases_md}

--- STRUCTURAL PATTERNS TO AVOID ---
{structures_md}

Rewrite any input text accordingly."""

Send system_prompt to Claude, OpenAI, or any LLM that respects system instructions to apply the Stop Slop rules dynamically.

Claude Skill Usage

Place the repository folder where Claude can access it, then reference the skill:

<skill>
  name: stop-slop
  path: /path/to/stop-slop
</skill>

With the skill loaded, Claude automatically reads SKILL.md, follows the links to references/phrases.md and references/structures.md, and applies the transformations. For example, given the input:

"Here's the thing: building products is hard. Not because the tech is complex, because people are complex."

Claude outputs:

"Building products is hard. Technology is manageable. People aren't."

See references/examples.md in the repository for additional before-and-after samples.

Command-Line Processing

For quick prototyping, download the files and apply phrase filtering with standard Unix tools:

curl -L https://raw.githubusercontent.com/hardikpandya/stop-slop/main/references/phrases.md -o phrases.md
curl -L https://raw.githubusercontent.com/hardikpandya/stop-slop/main/references/structures.md -o structures.md

# Basic phrase removal using sed (demo only)

sed -e '/Here’s the thing:/d' -e '/Full stop\./d' input.txt > cleaned.txt

While structural pattern matching requires a proper parser, the phrase list from phrases.md can be processed with simple text manipulation for rapid testing.

Summary

• Stop Slop reference files (phrases.md and structures.md) provide declarative, markdown-based rules for text cleanup.
• SKILL.md serves as the entry point that links to these reference files, creating a portable skill definition.
• The system uses a three-step lookup filter: identify matches, apply fixes, and iterate until clean.
• Being plain markdown, the files integrate with Python scripts, Claude skills, and command-line workflows without requiring code execution.

Frequently Asked Questions

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

phrases.md targets specific lexical items—filler phrases like "Here's the thing" or weak adverbs like "very"—and instructs the model to delete or replace them. structures.md targets syntactic patterns, such as binary contrasts or rhetorical negations, and instructs the model to rewrite the entire sentence structure into a direct statement.

How do I add custom phrases to Stop Slop?

Edit references/phrases.md or references/structures.md directly. Because the skill logic remains untouched in SKILL.md, you can add new rows to the markdown tables without modifying any code. The changes take effect the next time the skill is loaded, whether in Claude or a custom script that reads the files.

Can I use Stop Slop reference files with GPT-4 or other LLMs?

Yes. Since the reference files are plain text, you can concatenate phrases.md and structures.md into the system prompt of any LLM that follows instructions, including GPT-4, Gemini, or local models. The Python example above demonstrates fetching the files and formatting them for arbitrary APIs.

Where are the Stop Slop reference files located in the repository?

The files reside in the references/ directory at the root of the hardikpandya/stop-slop repository. references/phrases.md contains the banned phrase list, references/structures.md contains the structural patterns, and references/examples.md provides sample transformations for testing.

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