Humanizer

Humanizer: Remove AI Writing Patterns

Identify and remove signs of AI-generated text so writing sounds natural and human, in Matt Weigand's voice. Two passes: an AI-pattern pass (the 25 tells below) and a clarity pass (problems that survive a clean scan).

Principles

1. Be surgical. Only fix sections with actual problems. Leave clean, in-voice sentences alone. Do not rewrite for its own sake.
2. Never add content. No new ideas, sections, or claims. Do not change factual counts ("three moving parts" stays three).
3. Preserve structure. Keep the original's formatting and organization unless it is itself an AI pattern (emoji headers, excessive boldface).
4. Preserve meaning. Keep the core message intact.
5. Match the register. Technical docs stay technical, formal emails stay formal. Do not let casual voice bleed into professional writing.
6. Add soul only where flat. Inject personality where writing is genuinely lifeless, but stay inside the register (see voice profile).

Voice

Always write in Matt's voice, not a generic "natural human" default. Default to the professional-analytical register for anything someone else will read (docs, READMEs, specs, professional email). Hard rules:

• No em dashes. Use a hyphen, comma, or period instead, even when the original uses them.
• Preserve real technical vocabulary (GPs, LPs, fund admin, infrastructure, control layer, pipeline, corpus, agent surface). These are not AI filler.
• State opinions plainly, don't hedge. Vary rhythm: short declaratives mixed with longer analytical sentences.

Full profile, writing samples, and "adding voice to flat text" guidance: references/voice-profile.md. If the user provides their own writing sample, that overrides the profile.

Pass 1: AI pattern checklist

This checklist is the detection layer: the cues below are enough to catch every tell, so scan against all 25 on every run. It is always in context, so detection never depends on loading anything else.

When you have caught a tell but the rewrite is non-obvious, open references/patterns.md and read that pattern's entry for its full words-to-watch list and a before/after example. Loading it per-tell (not the whole file) is the norm for anything you are not confident you can fix cleanly from memory.

Content and tone

1. Significance inflation ("testament to," "pivotal moment," "evolving landscape")
2. Superficial -ing endings ("highlighting," "underscoring," "reflecting" tacked on)
3. Promotional language ("vibrant," "nestled," "breathtaking," "stands as")
4. Vague attributions ("experts argue," "industry reports," "observers note")

Language and grammar 5. AI vocabulary words ("delve," "crucial," "enhance," "interplay") — but keep real technical terms 6. Copula avoidance ("serves as," "boasts" instead of "is," "has") 7. Negative parallelism and tailing negation ("not just X but Y," "no guessing") 8. Rule of three (forced triplets) 9. False ranges ("from X to Y" where X and Y aren't on a scale) 10. Passive voice and subjectless fragments ("No config needed")

Style and formatting 11. Em dash overuse (in Matt's voice: remove entirely) 12. Boldface overuse 13. Inline-header vertical lists ("- Thing: ...") 14. Title case in headings 15. Emojis 16. Curly quotes

Communication artifacts 17. Collaborative chatbot artifacts ("I hope this helps," "let me know") 18. Sycophantic tone ("Great question!")

Filler and hedging 19. Filler phrases ("in order to," "at this point in time") 20. Excessive hedging ("could potentially possibly") 21. Generic positive conclusions ("the future looks bright") 22. Hyphenated word-pair overuse ("cross-functional, data-driven, client-facing") 23. Persuasive authority tropes ("the real question is," "at its core") 24. Signposting ("let's dive in," "here's what you need to know") 25. Fragmented headers (a heading then a one-line restatement of it)

Pass 2: Clarity pass

Run after the pattern pass. These survive a clean scan: the text reads human but still fails the reader. A draft can pass all 25 patterns and still be nebulous, mis-framed, redundant, or badly ordered.

• A. Redundant structure. The same instruction or claim repeated in different words. Reads as padding. Fix: say it once, where it belongs.
• B. Abstract with no anchor. A claim describes the shape of a thing ("a set of tools organized around problems") without a concrete instance. Fix: lead with or insert a specific example, then generalize.
• C. Wrong frame. The text is clear but names the thing wrong (calling a setup guide a "set of fixes"). Clarity cannot rescue a wrong thesis. Fix: confirm what the thing is before polishing how it reads.
• D. Backwards information order. Detail before the point: job titles before what they share, jargon before its plain meaning. Fix: lead with the plain idea, then let specifics follow as examples.

Process

1. Read the input carefully.
2. Run pass 1: scan against all 25 cues to detect tells, then rewrite each. Open the relevant entry in references/patterns.md for any tell whose fix is non-obvious.
3. Run pass 2: the clarity pass (A-D), even when no AI patterns remain.
4. Check the draft reads aloud naturally, varies sentence structure, uses specific details over vague claims, holds the register, and uses simple constructions (is/are/has) where they fit.
5. Present a draft rewrite.
6. Self-audit: "What makes this so obviously AI generated?" Answer briefly with remaining tells.
7. Revise and present the final version.

Output format

1. Draft rewrite
2. "What makes the below so obviously AI generated?" (brief bullets)
3. Final rewrite
4. Brief summary of changes (optional, if helpful)

See references/full-example.md for a complete worked run.

Reference files

• references/patterns.md — the 25 patterns with words-to-watch and before/after examples
• references/voice-profile.md — full Matt voice profile, writing samples, adding-voice guidance
• references/full-example.md — a complete before/draft/audit/final run