---
name: "humanizing_ai_writing"
description: "Rewrite AI-sounding prose into human prose. Fixes narrative structure before word choice, preserves the author's voice, scrubs density-scored markers, and never invents facts. Use when a draft reads like it was assembled by a language model, before publishing AI-assisted text, or when reviewing someone else's draft for machine tells."
---

# Humanizing AI Writing

## Purpose

Turn AI-drafted text into prose that reads like a person wrote it. The method is structure first, markers second, voice throughout. It assumes the draft's facts are the only facts available; adding detail is allowed only from the source text or the user, never invented.

This is not a detector-evasion trick. The goal is better writing: writing where each sentence makes a choice a machine would not.

## Workflow

### 1. Find the point

Read the draft and state in one sentence what it is actually trying to say. If you cannot, that is the first problem. Most AI drafts bury the point under a throat-clearing opener and restate it three times. Name the point before touching a word.

### 2. Restructure first

Do this before any word-level edit. Machine rhythm lives in the skeleton:

- Delete the throat-clearing opener. If the first paragraph could introduce any topic on earth, it introduces nothing.
- Delete the summary closer. If the last paragraph just repeats what was said, end when the argument is made.
- Break symmetric paragraph shapes. The tell is not one topic-sentence-to-summary paragraph; it is three in a row built the same way. Merge two paragraphs, split one, let the complexity of the idea set the length.
- Give uneven weight to uneven evidence. AI gives every point equal space. Humans dwell where the strong evidence is and compress the dull parts.
- Replace transition words with real connections. If the second paragraph does not need "furthermore," the relationship between the ideas does the work. A sentence that says something beats a sentence that signals something.
- Move the strongest sentence to the front. AI builds to the point; humans lead with it.

### 3. Preserve the voice

Match the author's register, not yours. If the author writes plainly, keep it plain. If they write formally, keep it formal, but cut the formality the author would never use. The "would I say this out loud to a friend" test is too strict for a legal memo and right for a blog post; calibrate it to the document, then apply it to every sentence.

### 4. Scrub markers by density

Run `bin/check.py` on the draft. Work from its output, but never mechanically. See `references/markers.md` for the full lexicon and the reasoning behind each budget:

- Em dashes over ~1 per 100 words. Replace with a comma, colon, parentheses, or a rewrite, never a period.
- Exactly-three lists past 4 per page, or triples used where two or four items would be natural. Adjective triples ("clear, concise, and compelling") get broken first.
- Hedging stacks ("might potentially be argued", "could possibly suggest"). One hedge is honesty; two is the model managing uncertainty.
- Vocabulary markers at 2-5x human rate: significant, crucial, notably, comprehensive, robust, leverage, landscape, nuanced, foster, streamline, elevate, empower, utilize, facilitate, harness, unlock, navigate, seamless, ecosystem. One per paragraph is fine; three is a signature. Prefer the plain word ("use" for "leverage" and "utilize"), a real number for "significant", deletion for "notably" and "multifaceted".
- Filler adverbs: fundamentally, essentially, ultimately, crucially, arguably, undoubtedly. Delete whole.
- "It's not X, it's Y" and "not just X" frames. Keep only when the negation is the actual content.
- Sycophancy and sincerity openers: "Great question", "I'd be happy to", "Let me be honest", "To be honest". Delete; start with the answer.
- Nominalizations and polysyllabic Latinate filler where a verb would do ("implementation of strategies" to "how to").

### 5. Facts are sacred

Every claim, name, number, date, quote, and citation in the rewrite must trace to the source draft or to information the user supplied. Swapping a vague claim for a specific one is allowed only when the specific comes from the source or the user. If a sentence needs a concrete detail to land and none exists, write the plain version without it, or ask for it. Opinions and reactions are voice, not facts, so adding stance is fine in voice-appropriate pieces. Never invent.

### 6. Eval loop

After the rewrite:

1. Run `bin/check.py` again. Every VIOLATION must be gone or consciously kept with a reason. REVIEW counts must drop.
2. Verify meaning: list the draft's factual claims and confirm each survives in the rewrite with the same meaning. Compressing is fine; changing is not.
3. Read it aloud (or silently, at speaking pace). Rewrite anything you would never say.
4. Over-correction check: compare the before and after. If the rewrite reads like a different kind of machine text (forced informality, inserted quirks, manufactured imperfections), undo the last round of changes. Edits stay proportional to real problems; when in doubt, leave it.

## Output Contract

- The rewrite, in the author's register, with the same factual claims as the source.
- A short change log: what was restructured, which markers were scrubbed (with before/after counts from the checker), and anything you cut because it could not be verified.
- Anything you could not fix (missing detail, unclear point) listed as open questions for the author, not silently patched.

## Operating Rules

1. Structure before words, always. A marker scrub on an unchanged skeleton still reads like AI.
2. Never invent facts. Unverifiable detail gets cut or flagged, never fabricated.
3. Density, not presence. One em dash or one "however" is writing. A budget overrun is the tell. The checker's budgets encode this; judgment handles the rest.
4. Preserve the author's voice; do not impose yours. Adding personality is allowed only where the piece and the author call for it.
5. No em dashes in the rewrite except the rare one the author would actually write, and never in the change log. No "furthermore", "moreover", "in conclusion". No stacked triple adjectives. No rhetorical-question setups.
6. The eval loop is not optional. Run the checker before and after, verify meaning survived, and check for over-correction.