Skip to main content

​
What is “Slop”?

“Slop” refers to low-quality LLM output patterns:
  • Excessive verbosity - 10 paragraphs when 2 sentences suffice
  • Unnecessary preambles - “Great question! I’d be happy to help…”
  • Filler content - Restating the question, obvious observations
  • Over-hedging - “It depends”, “There are many factors…”
  • Emoji abuse - Unprompted decorative emojis
  • Repetition - Saying the same thing multiple ways

​
Why It Matters

Slop wastes tokens, time, and attention: 
Sloppy response: 500 tokens, 2 useful sentences
Clean response:  50 tokens, same 2 sentences

Cost difference: 10x
Readability: Much worse
Actionability: Buried in noise

​
Constraint Techniques

​
Direct Instructions

Add anti-slop rules to prompts: 
## Output Rules
- No preambles or pleasantries
- No restating the question
- No "I'd be happy to help" or similar
- No emojis unless explicitly requested
- Max 3-5 sentences unless complexity requires more
- Lead with the answer, not context

​
Configuration by Tool

Add anti-slop rules to your AI tool’s config file:
File: CLAUDE.md (project root or ~/.claude/)
## Communication Preferences
- Be direct and concise (no unnecessary verbosity)
- Keep responses 25-40 lines max
- Only use emojis if explicitly requested
- No filler phrases ("Great question!", "I'd be happy to...")
- Lead with conclusions, support with evidence

​
Structured Output Requests

Force structure to prevent rambling: 
Respond in this exact format:
- **Answer**: [1-2 sentences]
- **Reason**: [1 sentence]
- **Next step**: [1 action item]

​
Anti-Patterns to Block

​
The Preamble Problem

# Bad
"Great question! I'd be happy to help you with that.
Let me think about this carefully. There are several
aspects to consider here..."

# Good
"The issue is X. Fix it by doing Y."

​
The Summary Trap

# Bad
"In summary, to summarize what we discussed,
the main points are: [repeats everything]"

# Good
[Just answer directly, no meta-commentary]

​
The Hedge Spiral

# Bad
"It depends on various factors. There are many
considerations. It's hard to say definitively.
Generally speaking, in most cases..."

# Good
"For your case: Do X. Exception: if Y, do Z instead."

​
Measurement

​
Token Efficiency Ratio

Efficiency = (Useful content tokens) / (Total tokens)

Target: > 80% efficiency
Red flag: < 50% efficiency

​
Quality Signals

Good SignsBad Signs
Answers firstPreambles first
Specific actionsVague suggestions
Concrete examplesAbstract concepts
Direct languageHedged language
Appropriate lengthAlways max length

​
Implementation

​
Agent Prompt Template

# Agent: [Name]

## Anti-Slop Rules (CRITICAL)
You MUST follow these output rules:
1. Never start with "I", "Great", "Sure", or similar
2. Never use emojis unless in user-facing content
3. Never restate what you were asked
4. Never summarize at the end
5. Max 5 sentences unless task requires more
6. Lead with the answer or action

## Task
[Task description]

​
Review Checklist

Before accepting agent output:
  • Starts with substance, not preamble
  • No filler phrases
  • No unnecessary hedging
  • No emoji spam
  • Appropriate length for content
  • No repetition or restating
  • No meta-commentary (“In this response…”)

​
Calibration

​
When Verbosity is Appropriate

Slop constraints don’t mean terse at all costs:
  • Teaching contexts - Explanations need detail
  • Critical decisions - Nuance matters
  • User-facing content - Personality is valuable
  • Complex topics - Brevity can mislead

​
Adjusting by Context

# For internal agent tasks
Be extremely concise. Output only what's needed.

# For user explanations
Be clear and helpful. Use examples. Explain reasoning.

# For documentation
Be thorough but organized. Use headers and structure.

​
Common Fixes

​
Fix Over-Qualification

# Before
"Based on my understanding, I believe the issue might
potentially be related to..."

# After
"The issue is [X]."

​
Fix Unnecessary Structure

# Before
"## Overview
Let me provide an overview...

## Analysis
Here's my analysis...

## Conclusion
In conclusion..."

# After
"The problem is X. Solution: Y."

​
Fix Explanation Creep

# Before
"To understand this, first we need to understand A,
which requires understanding B, which means C..."

# After
"Do X. (Background: A affects B.)"

​
Best Practices

  • Add anti-slop rules to all agent prompts
  • Review output token efficiency regularly
  • Use structured output formats
  • Train on examples of ideal responses
  • Measure and track verbosity metrics
Balance is key - Too aggressive anti-slop creates curt, unhelpful responses. Calibrate to context.

Token Economics

Reduce costs through efficiency

Context Optimization

Quality input = quality output