Humanize Writing

Removes the statistical fingerprints of LLM-generated text, as catalogued by Wikipedia's WikiProject AI Cleanup ("Signs of AI writing") and confirmed by corpus studies of 15M+ PubMed abstracts (Kobak et al. 2024, arXiv:2406.07016).

The core principle: no single pattern matters much in isolation. What exposes AI text is the density of patterns appearing together. Write so the cluster never forms.

Two modes

Generation mode (default). You are writing new prose. Apply the rules while drafting so the text comes out human the first time.

Rewrite mode. The user pastes existing text and asks you to clean it. Strip AI patterns while preserving everything else:

• Never alter direct quotes. Quoted third-party text stays verbatim, even if it contains banned patterns.
• Never change citations, numbers, statistics, dates, URLs, or proper nouns.
• Preserve the author's meaning, facts, and claims exactly. If removing a pattern would change what the text asserts, keep the assertion and only change the phrasing.
• Don't "improve" content they didn't ask you to change. Strip patterns; don't editorialize.

Enter rewrite mode when the user provides text to fix; otherwise stay in generation mode.

Detect the register first

A cold email, a LinkedIn post, a blog article, and a formal report should not be humanized identically. Before applying any rule, infer the register from the user's request and context:

| Register | Contractions | Fragments | Em dashes | Example contexts | |----------|--------------|-----------|-----------|------------------| | Social | yes | yes | rare | LinkedIn, X/Twitter, Slack, chat | | Email | yes | sparingly | rare | cold email, follow-ups, newsletters | | Editorial | yes | rarely | max 1/300 words | blog posts, articles, essays | | Formal | no | no | avoid | reports, proposals, formal docs, academic |

All other rules apply in every register. When the register is ambiguous, ask yourself who receives the text and default to the closest row.

Workflow

1. Detect mode (generation vs rewrite) and register.
2. If a voice-profile.md exists in the project, or the user offers samples of their own writing, read references/voice.md and apply their profile instead of the generic clean voice.
3. Draft (or rewrite) with the core rules below in mind.
4. Run the self-audit checklist against the draft.
5. Fix every violation, then deliver.

Read the reference files when you need depth:

• references/patterns.md — the full numbered catalog (36 patterns, each with a before/after pair). Read it in rewrite mode, or whenever the self-audit flags something and you need the precise fix.
• references/vocabulary.md — the banned-word list with plain replacements, grouped by category. Read it when substituting words.
• references/voice.md — voice calibration: extracting a profile from user samples and persisting it.

Core rules

The most damaging patterns, always in effect. Numbers refer to references/patterns.md.

Structure (P1–P12). Never write negative parallelism ("It's not X, it's Y" / "This isn't about speed. It's about trust.") — state the positive claim directly. Break the rule of three: one strong word or two, not "innovative, efficient, and scalable". No false ranges ("from startups to enterprises"). No trailing participles that editorialize ("...cementing its status as a leader") — state the fact and stop. Use plain "is/are" instead of "serves as", "stands as", "boasts". Call a thing the same name twice instead of cycling synonyms. No staccato drama fragments ("One goal. Zero excuses."). Collapse hedging stacks ("could potentially help" → "may help" or commit). Cut filler ("in order to" → "to").

Framing (P13–P20). No summary closers ("In conclusion", "Overall", a final paragraph restating the piece) — end on your last substantive point. Cut "It's important to note", "Notably", "Interestingly". Name your sources or cut the claim — no "experts say", "studies show", "widely regarded as". No significance inflation ("pivotal moment", "enduring legacy"). No "Despite challenges... the future looks bright" scaffolds. Don't open with a definition or a restatement of the prompt — open with the most interesting true thing.

Conversational artifacts (P21–P26). Strip anything a chatbot says to its user: "Great question!", "Honestly?", "I hope this helps!", "Would you like me to...", "As of my last update...", and unfilled placeholders like "Insert name]". These must never appear inside a deliverable.

Formatting (P27–P34). At most one em dash per ~300 words, and only where a comma or parentheses genuinely wouldn't work. No bold mid-sentence for emphasis. No "Term: definition" bullets. Default to prose — bullets only when the user asks or the content is truly enumerable. No emoji in headings, ever. No headings at all under ~400 words. Sentence case for headings. Keep quotation marks consistent (don't mix curly and straight).

Vocabulary (P35–P36). Avoid the AI word list — delve, tapestry, intricate, pivotal, crucial, underscore, landscape (metaphorical), foster, testament, boast, meticulous, realm, showcase, leverage (verb), robust, seamless, elevate, embark, journey (metaphorical), navigate (metaphorical), unlock, harness, empower, game-changer, cutting-edge, groundbreaking, transformative, comprehensive, holistic, streamline, synergy, paradigm, myriad, plethora, vibrant, ever-evolving, "in today's fast-paced world", "at the end of the day", deep dive — unless the user's own text uses them or no plain alternative exists. Full list and replacements in references/vocabulary.md.

Rhythm. Vary sentence length deliberately; follow a long sentence with a short one. Prefer concrete specifics ("replies within two hours") over abstractions ("prompt communication"). Commit to claims at the scale they're actually true. Say each idea once — cut the echoes.

Self-audit checklist

Before delivering any prose, scan the draft for:

1. Any "not X, but Y" / "isn't just X — it's Y" construction (P1–P2) → rewrite as a direct claim
2. Em dashes — more than 1 per 300 words (P27)? → replace with commas or parentheses
3. Any banned-vocabulary word (P35) → replace with the plain alternative
4. Triplet lists (P3) → cut to one or two, or expand honestly
5. Final paragraph (P13) → does it summarize instead of end? Replace with a specific closing point
6. Trailing "-ing" significance clauses (P5) → delete
7. "It's important to note" and cousins (P14), chatbot phrases (P21–P26) → delete
8. Bold mid-prose, "Term: definition" bullets, emoji headings (P28–P31) → strip
9. Unnamed "experts/studies" (P15) → name them or cut
10. Read three consecutive sentences aloud — same length and rhythm? → break one up

Optional mechanical check: run scripts/ai_pattern_lint.py on the draft; it reports pattern hits per 1000 words.

False-positive guard

These patterns are statistical signals, not proof of anything. Apply them with judgment:

• The user wins. If they explicitly ask for bullets, em dashes, bold terms, or any banned element, their instruction overrides this skill. It sets defaults, not handcuffs.
• Quotes are untouchable. Never "fix" quoted third-party text, in either mode.
• Precision beats style. Never sacrifice factual accuracy, a necessary technical term, or a legally required phrase to dodge a pattern. If "comprehensive" is the accurate word, keep it.
• Humans use these too. An em dash or a three-item list is not a crime; the density of many patterns together is the tell. Fix the cluster, not every isolated instance.