DEV Community

Cover image for One Open Source Project a Day (No. 78): stop-slop - A Skill File That Teaches AI to Eliminate Its Own Writing Tells
WonderLab
WonderLab

Posted on

One Open Source Project a Day (No. 78): stop-slop - A Skill File That Teaches AI to Eliminate Its Own Writing Tells

#opensource #ai #claude #promptengineering

Introduction

"If it sounds like a pull-quote, rewrite it."

This is the 78th article in the "One Open Source Project a Day" series. Today's project is stop-slop.

You know the writing. It opens with "In today's fast-paced world..." The second paragraph has "it's worth noting that." The middle is full of "undeniably," "transformative impact," "revolutionary breakthrough." It ends with a profound one-liner. You recognize it instantly — this was written by an AI.

These patterns have become the signature tells of AI-generated content. stop-slop is a Claude Code Skill file with a simple mission: teach AI to recognize and actively eliminate its own writing tells. 5.8k Stars, pure Markdown, no code, works in any AI tool that supports custom system prompts.

What You Will Learn

  • The 8 categories of AI writing patterns that give it away
  • How the 5-dimension scoring rubric quantifies whether text "sounds human"
  • The core design insight: prevent patterns at generation time rather than detect them afterward
  • How to integrate these rules into your own writing workflow
  • Why this project spawned a significant community of derivative projects

Prerequisites

  • Experience using Claude or other LLMs for writing assistance
  • An intuitive sense of "AI-flavored writing" — that feeling when something reads off

Project Background

Project Introduction

stop-slop was created by product designer Hardik Pandya. He's a serious writer who, after deep use of AI-assisted writing, noticed a problem: AI-generated text follows a highly predictable set of patterns — regardless of topic, regardless of prompt, the same opening styles appear, the same pivot structures, the same "grand narrative" tone.

His solution wasn't "make AI write better in some abstract sense." It was to explicitly enumerate all the known AI writing tells, then tell the AI: when you see these, fix them.

The execution is rigorous — 8 rules, 3 reference files, a pre-publish checklist, a scoring rubric — forming a complete system that can be embedded into any writing workflow.

Author

  • Author: Hardik Pandya (hardik.substack.com)
  • Background: Product designer, longtime writer
  • Distribution: GitHub + Substack (with detailed usage notes and background essay)

Project Data

  • ⭐ GitHub Stars: 5,800+
  • 🍴 Forks: 435+ (spawned multiple derivative projects)
  • 📄 License: MIT
  • 📁 Format: Pure Markdown, zero code
  • 🔌 Compatible with: Claude Code, Claude Projects, custom instructions, direct API system prompts
  • 🌐 Repository: hardikpandya/stop-slop

Main Features

Core Utility

stop-slop is a Skill file. Once installed, it activates automatically during writing or editing tasks and systematically identifies and removes AI writing tells.

Your prose draft
      ↓
stop-slop Skill (activates)
      ↓
  Identifies 8 categories of AI writing patterns
      ↓
  Corrects each one
      ↓
  5-dimension score (< 35/50 = needs revision)
      ↓
Output: writing that sounds like a human wrote it
Enter fullscreen mode Exit fullscreen mode

Installation and Usage

Claude Code (Skill directory install):

# Clone into Claude's global skills directory
git clone https://github.com/hardikpandya/stop-slop.git ~/.claude/skills/stop-slop

# After installation, the Skill activates automatically during writing tasks
Enter fullscreen mode Exit fullscreen mode

Claude Projects / Custom Instructions:

# Copy the contents of SKILL.md into "Custom Instructions" or "System Prompt"
Enter fullscreen mode Exit fullscreen mode

Direct API:

import anthropic

with open("SKILL.md", "r") as f:
    skill_content = f.read()

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-4-6",
    system=skill_content,
    messages=[{"role": "user", "content": "Please edit this passage: ..."}]
)
Enter fullscreen mode Exit fullscreen mode

Deep Dive

The 8 Core Rules

The heart of SKILL.md. Each rule targets a specific category of AI writing patterns:

Rule 1: Cut Filler

Targets: throat-clearing openers, emphasis crutches, all adverbs

Typical AI tells:
  ✗ "In today's rapidly evolving digital landscape..."
  ✗ "It's worth noting that"
  ✗ "Certainly," "Indeed," "Obviously," "Very," "Extremely"
  ✗ "It goes without saying"

Fix: Start with the actual point. Remove all adverbs.
Enter fullscreen mode Exit fullscreen mode

Rule 2: Break Formulaic Structures

Targets: binary contrasts, dramatic fragments, rhetorical setups, false agency

Typical AI tells:
  ✗ "It's not about X, it's about Y" (binary contrast)
  ✗ "But here's the thing." (dramatic standalone sentence)
  ✗ "So how do we...?" (rhetorical question setup)
  ✗ "This opens new possibilities" (false agency)

Fix: Just say Y. Cut the setup. Name who is doing what specifically.
Enter fullscreen mode Exit fullscreen mode

Rule 3: Active Voice

Every sentence needs a human subject performing an action.

  ✗ "This feature was designed to..."
  ✗ "Research suggests..." (research with no author)
  ✓ "The team designed this feature to..."
  ✓ "MIT researchers found..."
Enter fullscreen mode Exit fullscreen mode

Rule 4: Be Specific

  ✗ "This is a watershed moment"
  ✗ "Every single person..." / "Always" / "Never" (lazy absolutes)
  ✓ "[Specific event] marked [specific change]"
Enter fullscreen mode Exit fullscreen mode

Rule 5: Put the Reader in the Room

  ✗ "People often find themselves..." (narrator-from-a-distance)
  ✓ "You'll find..." (direct address)

  "You" beats "people." Specific scene beats abstract description.
Enter fullscreen mode Exit fullscreen mode

Rule 6: Vary Rhythm

  ✗ Three consecutive sentences of the same length
  ✗ Em dashes — used like this — for emphasis
  ✗ Things come in threes: first..., second..., third... (always three)
  ✓ Mix long and short sentences. Two items are stronger than three.
Enter fullscreen mode Exit fullscreen mode

Rule 7: Trust Readers

State facts directly. Skip softening and hand-holding.

  ✗ "You might be wondering..."
  ✗ "Before we dive in, let me explain..."
  ✓ Give the information. Readers will judge it themselves.
Enter fullscreen mode Exit fullscreen mode

Rule 8: Cut Quotables

"If it sounds like it was written to be screenshot-shared, rewrite it."

  ✗ "Technology is the tool. People are the core."
  ✗ "True innovation begins with redefining the problem."

These sentences usually say nothing — they sound like they mean something
but don't actually contain information.
Enter fullscreen mode Exit fullscreen mode

Pre-Publish Checklist

SKILL.md includes a 12-item checklist — the most practical part for daily use:

Check Action
Any adverbs? Remove them all
Passive voice? Find the actor, make them the subject
Inanimate subject doing a human action? Name the person
Sentence starts with a Wh- word? Restructure
"Here's what/this/that" opener? Cut to the point
"Not X, it's Y" contrast? Just say Y
Three consecutive same-length sentences? Break one
Punchy one-liner at paragraph end? Vary it
Em-dash present? Delete
Vague declarative? Name the specific thing
Narrator-from-a-distance voice? Place the reader in the scene
Meta-joiners like "the rest of this essay"? Delete; let the writing move

The 5-Dimension Scoring Rubric

Each dimension scores 1–10, maximum 50. Below 35/50: revise.

Dimension Core Question
Directness Making statements, or making announcements?
Rhythm Varied sentence lengths, or metronomic regularity?
Trust Respects reader intelligence, or hand-holding?
Authenticity Sounds human, or sounds AI-generated?
Density Every word earning its place, or cuttable content present?

The rubric's value: it turns "sounds like AI" from a vague intuition into five specific, improvable dimensions. When you're revising, you know what you're fixing. You also know when you can stop.

Prevention, Not Detection

This is stop-slop's most important design insight, and what separates it from most "AI content detection" tools:

The mainstream approach (detection):
  AI-generated content → detector → flagged as "AI-written" → ...now what?

stop-slop's approach (prevention):
  Before/during writing → Skill activates → AI avoids patterns while generating
  → Clean output, directly
Enter fullscreen mode Exit fullscreen mode

The difference is fundamental. Detection tools tell you "this was written by AI" but don't help you fix it. stop-slop is real-time interception during the writing process, not post-hoc forensics.

Community Derivatives

Behind the 5.8k Stars, the more interesting number is 435 Forks — and the derivative projects they spawned:

  • skill-deslop: Customized for scientific writing, targeting patterns specific to academic papers
  • anti-ai-slop-writing: An enhanced version targeting "statistically detectable AI writing patterns," covering more edge cases

These derivatives tell you something: the need to identify and eliminate AI writing tells is real, persistent, and domain-specific — every field has its own particular flavor of slop.

Project Links & Resources

Official Resources

Target Audience

  • Content creators: Using AI to help with blog posts, articles, and social content — want the output to sound like themselves
  • Technical writers: Writing docs, tutorials, API references — don't want "AI flavor" undermining readability
  • Brand and marketing teams: AI-generating content at scale, need consistent brand voice
  • Anyone using AI for writing: Internalize these rules and they'll improve your own writing too

Summary

Key Takeaways

  1. 8 concrete rules: Each targets a specific AI writing tell, from filler phrases to false agency — covering the most common patterns
  2. 5-dimension scoring rubric: Turns "sounds like AI" into five improvable dimensions; 35/50 is the passing line
  3. Prevention over detection: Intercepts during writing rather than flagging afterward — a fundamental difference in approach
  4. Zero code, universal: Pure Markdown, any tool supporting custom system prompts can use it directly
  5. Community traction: 435 forks and multiple domain-specific derivatives confirm this problem is widespread

One-Line Review

stop-slop did something that looks simple but isn't: it decomposed the vague intuition of "AI-flavored writing" into 8 actionable, specific rules — and with concrete rules, AI can self-correct instead of hoping for the best.

Find more useful knowledge and interesting products on my Homepage

Top comments (0)