Skip to content

phjlljp/humanize

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Humanize

Make AI text sound human.

Humanize is a Claude Code plugin that transforms AI-generated writing into text that reads like a person wrote it. It works by detecting and eliminating 43 documented AI writing anti-patterns drawn from Wikipedia's "Signs of AI writing" and Grokipedia's analysis of AI-generated text.

The patterns span 7 categories (vocabulary, syntax, content, structure, formatting, behavioral tells, and statistical signals) and are ranked into 3 severity tiers. The plugin runs a 4-phase process: detect patterns in the input, rewrite to remove them, audit the rewrite for survivors, then output the result.

Installation

Add the marketplace and install the plugin:

/plugin marketplace add phjlljp/humanize
/plugin install humanize@phjlljp

Usage

/humanize <text or file path>

Pass inline text or a path to a file. Humanize will scan for AI patterns, rewrite, audit, and return the cleaned text.

Examples

Basic transformation:

/humanize "The company's innovative approach serves as a testament to their commitment to excellence, showcasing the transformative potential of cutting-edge technology."

Transform with a diff showing what changed and why:

/humanize --diff "Additionally, the platform offers a comprehensive suite of tools designed to enhance productivity and foster collaboration."

Casual register with thorough cleanup (all tiers):

/humanize --casual --thorough ./draft-blog-post.md

Academic register:

/humanize --academic ./paper-intro.md

Flags

Flag Description
--diff Show a before/after comparison with pattern IDs explaining each change
--casual Target conversational register: contractions, shorter sentences, first/second person
--formal Target formal register: precise vocabulary, no contractions, third person
--academic Target academic register: passive voice where conventional, technical precision, reduced hedging
--thorough Address all 3 tiers including subtle patterns (default handles Tier 1 and Tier 2 only)

When no register flag is provided, Humanize auto-detects the register of the input and matches it.

How it works

  1. Detection & analysis -- Scan input against all 43 patterns. Produce a diagnostic report listing every detected pattern with its ID, severity tier, and the specific text that triggered it.
  2. Transformation -- Rewrite the text to eliminate detected patterns. Tier 1 (dead giveaways) and Tier 2 (strong signals) are always addressed. Tier 3 (subtle signals) only with --thorough.
  3. Audit -- Re-scan the transformed text. Ask: "If I read this cold, would I suspect AI wrote it?" Fix anything that still feels generated.
  4. Output -- Present the final text. If --diff was passed, include a paragraph-by-paragraph comparison showing what changed and which pattern motivated each change.

Pattern catalog

43 patterns across 7 categories. Tier 1 patterns are dead giveaways that immediately signal AI origin. Tier 2 patterns are strong signals that build a cumulative AI impression. Tier 3 patterns are subtle and only targeted with --thorough.

A: Vocabulary and word choice

ID Pattern Tier
A1 Overused AI vocabulary (delve, pivotal, tapestry, etc.) 1
A2 Overused multi-word phrases (shed light on, at its core, etc.) 1
A3 Excessive formal connectors (Additionally, Moreover, etc.) 1
A4 Positive intensifiers/buzzwords (innovative, cutting-edge, etc.) 2
A5 Copula avoidance (serves as, stands as, features, etc.) 2
A6 Elegant variation / synonym cycling 2

B: Syntactic patterns

ID Pattern Tier
B1 Negative parallelisms (not only...but also) 2
B2 Rule of three (adjective, adjective, and adjective) 2
B3 Superficial -ing analyses (highlighting the significance...) 1
B4 False ranges (from X to Y with no real scale) 2
B5 Repetitive syntactic templates 3
B6 Uniform sentence length 3
B7 Excessive nominalization 3

C: Content and semantic

ID Pattern Tier
C1 Exaggerated significance (pivotal moment, testament to) 1
C2 Promotional/advertisement tone (vibrant, nestled, breathtaking) 2
C3 Vague attributions/weasel words (experts argue, widely regarded) 2
C4 Overgeneralization (sweeping claims without evidence) 2
C5 Superficial/generic analysis 2
C6 Undue emphasis on notability/media coverage 2
C7 Sentiment homogeneity 3

D: Structural

ID Pattern Tier
D1 Formulaic conclusions (In conclusion, the future holds promise) 1
D2 Challenges-and-recovery formula (Despite its...faces challenges) 1
D3 Inline-header vertical lists (bolded headers with colons) 2
D4 Rigid predictable organization 3
D5 Section summaries (In summary, to recap) 2
D6 Title-as-proper-noun leads 2
D7 Vague see-also/related sections 3

E: Formatting and style

ID Pattern Tier
E1 Overuse of boldface 2
E2 Title case in headings 2
E3 Em dash overuse 2
E4 Emoji decoration 2
E5 Curly quotation marks (inconsistent) 3
E6 Unnecessary tables 3
E7 Non-standard heading styles 3

F: Behavioral/conversational

ID Pattern Tier
F1 Collaborative communication (I hope this helps, let me know) 1
F2 Knowledge-cutoff disclaimers (as of [date]) 1
F3 Sycophantic/servile tone (Great question!) 1
F4 Inappropriate direct address (you can, did you know) 2
F5 Phrasal templates/placeholders ([Your Name Here]) 1
F6 Didactic disclaimers (it's important to note) 2

G: Statistical/quantitative

ID Pattern Tier
G1 Low lexical diversity 3
G2 Reduced sentence length variation 3
G3 Excessive hedging (could potentially possibly) 2

Sources

Wikipedia: Signs of AI writing, maintained by WikiProject AI Cleanup

Grokipedia: Signs of AI-generated writing

Credits

The idea for this plugin came from the Humanizer skill by blader. The "Personality and Soul" approach, that removing AI patterns is only half the job and that clean but voiceless writing is just as suspect, is also directly adapted from that project.

License

MIT

About

Make AI text sound human. Claude Code plugin that detects and eliminates 43 AI writing anti-patterns.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors