Antislop

The AntiSlop

A comprehensive AI writing pattern detector and fixer. Combines patterns from Wikipedia's Signs of AI Writing with advanced structural detection and an editor mode that actually fixes problems.

The 30-Second Test

The Horoscope Test:

"Could anyone have written this, for anyone?"

If yes, it's slop. Like a horoscope -- technically applicable to everyone, resonant with no one.

What fails:

• Vague claims without specific examples
• Advice that applies universally without context
• Content missing the author's distinct perspective
• Writing that could have any byline

What passes:

• Specific tools, dates, outcomes mentioned
• Personal observations grounded in experience
• Opinions that not everyone would agree with
• Details only this author would know

Usage

/antislop

[paste your text here]

Or ask Claude to check text directly:

Please run antislop on this: [your text]

How It Works

1. Run the Horoscope Test - Could anyone have written this for anyone?
2. Scan for patterns - 50+ known AI tells across 6 categories, structure weighted above vocabulary
3. Calculate slop score - Tiered severity with quantifiable scoring
4. Apply fixes - Editor mode rewrites problems, not just flags them
5. Report changes - Before/after for every fix applied

Detection Patterns (50+, structure-weighted)

Tier 1: AI-vocabulary density (weight the cluster, not the word)

These words were near-certain AI tells in the 2023–2024 GPT-3.5/4 era. Frontier models in 2025–2026 emit the loudest of them far less, so a single hit now means little — treat this as a density probe, not a blocklist. Three or more clustered in a passage is still a reliable tell; one "leverage" in isolation usually is not. Score the density (see Scoring), not each word.

Legacy signals — weakened, still worth flagging in clusters:

Current signals — the vocabulary today's models still lean on:

Research evidence (lexical tells peaked 2023–2024; density-based detection ages better than any word list):

• Finnish study (56,878 essays): "delve" usage increased 10.45x post-ChatGPT
• Georgia Tech (168.3M articles): "delve" went from 0.31 to 7.9 per 1,000 papers in Q1 2024
• Biomedical study: co-usage of "delve," "realm," "underscore" increased up to 85x in 2023-2024
• 2025–2026 shift: frontier models were tuned against the loudest 2023 tells, so detection now leans on structure (below), where the signal has held.

Tier 2: Suspicious When Repeated

Problematic when overused or clustered.

Tier 3: Watch for Clusters

Fine individually, problematic together.

Content Patterns

Fabricated precision is a current-model tell: a suspiciously exact figure (47%, 3.2x, "over 10,000 hours") presented without a citation, to manufacture authority. Flag any hard statistic that has no source attached.

Language Patterns

Style Patterns

Structural Patterns (Critical)

These bypass phrase-based detection but are major tells.

Staccato Fragment Spam

Three or more consecutive short declarative sentences stating facts in parallel structure. AI's version of bullets pretending to be prose.

Before:

The model is impressive. Complex code ships fast. Documentation writes itself. Problems get solved quickly.

After:

The model is impressive -- complex code ships in a single session, documentation practically writes itself, and problems that would have taken a weekend now take an afternoon.

Detection rule: 3+ consecutive sentences that are all under 10 words, all declarative, following parallel structure, and could be bullet points.

Sentence Uniformity

Every sentence 10-15 words. Short. Punchy. Exhausting.

Real writing has rhythm -- mix 5-word sentences for impact with 25-word sentences that explore implications.

Comparator Sentences

Before:

This isn't theoretical. It's practical. This isn't a feature. It's a philosophy. It's not about X. It's about Y.

After:

Here's how it works in practice: [Just state what it is]

AI loves this rhetorical pattern. It sounds punchy but wastes words telling you what something isn't.

Over-Balanced Sections

Every section same length. All paragraphs 3-4 sentences. AI doesn't have opinions, so it gives balanced coverage to everything. Real writing reflects priorities.

Communication Patterns

Advanced Structural Tells

Manufactured Personality

AI trying to sound human but coming across as performative:

Before:

Five services. Five tabs. Five headaches. That got old fast. So I built an MCP server that unifies all of them.

After:

I use five different services for my workflow. They're solid platforms, but like most SaaS tools, each means another dashboard, another set of menus to navigate, another context switch.

No manufactured punch. No snark. Just describes the situation.

Self-Promotional Framing

Content positioning author's accomplishments as the headline instead of reader's transformation.

Before:

I shipped 11 projects over the holidays. Here's what I learned.

After:

Most developers aren't aware that [observation about the reader's situation]. Here's what's changing...

The author's experience is evidence, not the story.

Explanatory Header Templates

Headers that promise insight but deliver template structure:

• "Why This Actually Works"
• "What This Means For You"
• "The Real Reason..."
• "Here's What's Really Going On"

Fix: Replace with descriptive headers that summarize the actual content.

Engagement-Bait Openers

Rhetorical hooks that fake a conversation to manufacture interest:

• "What if I told you..."
• "Ever wondered why...?"
• "Let's be honest:"
• "Picture this:"
• "Here's the thing —"

Fix: Open with the actual claim or the specific detail. Trust the reader to keep reading because the content is worth it.

Trailing Affirmations

A short reassurance tacked onto a sentence to soften or self-approve:

Before:

Automation frees people to do what they do best. And that's a good thing. The tool handles the boring parts — which is exactly what you want.

After:

Automation frees people from the boring parts.

The affirmation adds no information. Cut it.

Filler and Hedging

Scoring System

Structure is weighted above vocabulary: modern models were tuned against the loudest lexical tells, so a clean-sounding passage can still be structurally slop.

Score interpretation:

• 0-5: Low risk (minor edits)
• 6-12: Medium risk (significant editing required)
• 13+: High risk (likely unedited AI output)

Editor Mode (Default)

This skill is an editor, not a critic. After detection:

1. Apply all fixes directly using the Edit tool
2. Report changes made with before/after examples
3. Save the cleaned file in place

Fix priority:

1. Remove all Tier 1 phrases
2. Deduplicate Tier 2 phrases (keep first, vary subsequent)
3. Break up staccato fragments (combine with em-dashes, commas, conjunctions)
4. Fix comparator sentences (just state what it is)
5. Vary sentence lengths where uniformity detected

To audit without editing, explicitly request "audit only."

Output Format

## AntiSlop Report

**Horoscope Test:** [PASS/FAIL] - [reason]
**Slop Score:** [X] → [Y] - [Risk Level]

### Fixes Applied

| Location | Before | After |
|----------|--------|-------|
| Line 3 | "Let's delve into the details" | "Here are the details" |
| Line 15 | "Game-changing approach" | "Different approach" |

### Remaining Considerations
- [Any issues requiring human judgment]

### The Core Principle
Your voice is in the specificity, the opinions, the rough edges, and the rhythm. Protect those.

Full Example

Before (AI-sounding):

Great question! Here is an essay on this topic. I hope this helps!

AI-assisted coding serves as an enduring testament to the transformative potential of large language models, marking a pivotal moment in the evolution of software development. In today's rapidly evolving technological landscape, these groundbreaking tools--nestled at the intersection of research and practice--are reshaping how engineers ideate, iterate, and deliver.

At its core, the value proposition is clear: streamlining processes, enhancing collaboration, and fostering alignment. It's not just about autocomplete; it's about unlocking creativity at scale.

• Speed: Code generation is significantly faster.
• Quality: Output quality has been enhanced.
• Adoption: Usage continues to grow.

In conclusion, the future looks bright. Let me know if you'd like me to expand!

After (Fixed):

AI coding assistants speed up some tasks. In a 2024 study by Google, developers using Codex completed simple functions 55% faster than a control group, but showed no improvement on debugging or architectural decisions.

The tools are good at boilerplate: config files, test scaffolding, repetitive refactors. They're bad at knowing when they're wrong.

Mira, an engineer at a fintech startup, said she uses Copilot "like autocomplete for boring code" but reviews every line before committing. Jake, a senior dev at a larger company, disabled it after it kept suggesting patterns from a deprecated internal library.

The productivity claims are hard to verify. GitHub says Copilot users "accept 30% of suggestions," but acceptance is not correctness.

Pattern Refresh Protocol

Patterns go stale as AI models evolve. Before scanning, check last-refreshed in frontmatter. If >30 days old, refresh first.

Refresh workflow:

1. Fetch the latest patterns from Wikipedia:

# Signs of AI writing - full wikitext
curl -s "https://en.wikipedia.org/w/api.php?action=parse&page=Wikipedia:Signs_of_AI_writing&prop=wikitext&format=json" | python3 -c "
import json, sys
data = json.load(sys.stdin)
print(data['parse']['wikitext']['*'][:30000])
" > /tmp/antislop-signs.txt

# WikiProject AI Cleanup
curl -s "https://en.wikipedia.org/w/api.php?action=parse&page=Wikipedia:WikiProject_AI_Cleanup&prop=wikitext&format=json" | python3 -c "
import json, sys
data = json.load(sys.stdin)
print(data['parse']['wikitext']['*'][:30000])
" > /tmp/antislop-cleanup.txt

2. Read the output and diff against patterns already in this skill
3. For genuinely new patterns not already covered:
   • Classify into Tier 1/2/3 based on how strongly they signal AI
   • Add to the appropriate table with example and fix
   • Update the pattern count in the overview
4. Update last-refreshed date in frontmatter
5. Report what was added (if anything)

Don't add duplicates. Many Wikipedia patterns are already covered here under different names. Only add patterns that represent genuinely new detection signals.

References

• Wikipedia: Signs of AI writing
• WikiProject AI Cleanup
• Finnish study on "delve" usage (56,878 essays)
• Georgia Tech analysis (168.3M articles)

Core Principle

AI slop isn't about individual words -- it's about patterns.

One "moreover" doesn't make content AI-generated. But "moreover" + "it's worth noting" + "delve into" + uniform sentences + emoji headers = obvious slop.

The goal is writing that sounds like a specific human with specific opinions, not a very polite committee trying not to offend anyone.