Writing styleguide
Antislop-informed rules for keeping AI-assisted writing honest
February 2026
This is an internal reference for writing on this site. It combines two sources: empirical data from the Antislop paper (Paech et al., 2025) on statistically overused patterns in LLM output, and the voice patterns documented in our own voice analysis. One tells us what to avoid. The other tells us what to aim for.
The goal isn't sterile writing. It's aware writing. Know the patterns. Decide deliberately.
Part 1: What to watch for antislop
The Antislop paper analyzed creative writing output from 67 language models against human baselines (Reddit creative writing + Project Gutenberg). They measured how often specific words and phrases appear in LLM text relative to human text. The ratios are staggering — some patterns are over 1,000x more frequent in model output.
These aren't bad words. They're overused words. A human writer might use "flickered" once in a story. A language model reaches for it reflexively, across every story, because alignment training reinforced it as a safe, literary-sounding choice.
Word-level banlist
Top overrepresented words across 67 models. The "% models" column shows how many models have this word in their top 120 most overrepresented words relative to human writing.
| word | % models | severity |
|---|---|---|
| flickered / flicker / flickering | 98.5 / 94.0 / 92.5 | high |
| leaned | 82.1 | high |
| muttered | 82.1 | high |
| gaze | 80.6 | high |
| grinned | 80.6 | high |
| gestured | 77.6 | high |
| murmured | 73.1 | medium |
| nodded | 73.1 | medium |
| glint | 68.7 | medium |
| hesitated | 68.7 | medium |
| whispered | 68.7 | medium |
| blinked | 64.2 | medium |
| hummed | 64.2 | medium |
| faintly | 62.7 | medium |
| unreadable | 62.7 | medium |
| shimmered | — | high (2,882x in gemma-3-12b) |
| stammered | — | high (3,833x in gemma-3-12b) |
| unsettlingly | — | high (3,833x in gemma-3-12b) |
| containment | 77.6 | medium |
| addendum | 74.6 | medium |
Context matters. "Flickered" in a story about a candle is fine. "Flickered" as the default verb for any light source in every story is slop. The question is always: did I choose this word, or did the model reach for its comfort zone?
Trigram banlist
Multi-word phrases that appear disproportionately in LLM output. These are the phrasings that make readers think "this sounds AI-generated" before they can articulate why.
| phrase | % models |
|---|---|
| voice barely whisper | 68.7 |
| said voice low | 61.2 |
| air thick scent / air thick smell | 49.3 / 19.4 |
| took deep breath / take deep breath | 44.8 / 32.8 |
| smile playing lips | 43.3 |
| something else something / something else entirely | 37.3 / 26.9 |
| could shake feeling | 31.3 |
| eyes never leaving | 29.9 |
| casting long shadows / long shadows across | 28.4 / 19.4 |
| heart pounding chest | 25.4 |
| spreading across face | 22.4 |
Structural patterns
Beyond individual words and phrases, models overuse certain sentence constructions.
Part 2: What to aim for voice analysis
From the voice analysis — patterns extracted from five years of raw journal entries. These are the things that make writing sound like a person, not a model.
Rhythm
Short punch. Short punch. Longer flowing thought that develops the idea. Short punch.
Tone
Transformation rules
When going from draft to published:
Preserve
- Fragment punches
- Profanity when earned
- Self-questioning
- The earnest plea ("Please.")
- Parenthetical asides
- Meta-awareness
- Specific numbers and timestamps
Avoid
- Over-explaining (trust the reader)
- Softening profanity
- Removing questions
- Excessive formality
- AI superlatives ("Absolutely!", "Great question!")
- Hedging ("It could be argued...")
- Bullet-point spam
Part 3: The editing pass synthesis
Where the two sources meet. Antislop tells us what patterns to scan for. The voice analysis tells us what to replace them with. Neither is sufficient alone — you can strip all the slop words and still end up with generic text if you don't know what voice to aim for.
Editing checklist
- Word scan — Search for banlist words. For each hit: did I choose this, or did the model? If the model, replace or cut.
- Phrase scan — Search for trigram patterns. These are harder to spot because they sound "literary." That's the trap.
- Structure check — Count "It's not X, it's Y" constructions. One per article might be fine. Three means the model was driving.
- Rhythm check — Read it aloud. Does it follow staccato-reflection-staccato? Or does every paragraph drone at the same length?
- Voice check — Does this sound like a person wrote it? Would you say this out loud? If it sounds like a LinkedIn post, rewrite.
- Hedging check — Cut "perhaps", "it could be argued", "one might say". State things or question them. Don't hedge.
The tension
Over-policing kills voice too. The antislop paper itself makes this point — hard bans on vocabulary cause worse problems than the slop they prevent. Their solution was "soft bans": reduce the probability of overused patterns without eliminating them entirely.
Same principle applies to manual editing. If "whispered" is genuinely the right word, use it. The goal is to notice when the model is on autopilot, not to avoid every word on a list. Awareness over avoidance.