Humanize

Purpose

Edit text to remove the specific, documented tells of AI-generated writing while preserving its meaning, coverage, and the author's voice. The tells are not vibes; they are catalogued patterns (grounded in Wikipedia's "Signs of AI writing", maintained by WikiProject AI Cleanup) that arise because models gravitate toward the most statistically likely phrasing that fits the widest variety of cases. The deliverable is a final rewrite produced through a draft, self-audit, final loop, never a single pass.

Hard rules

• Rewrite, do not delete. Every substantive point in the original survives into the rewrite; losing content is not humanizing, it is shrinking. Pure ceremony is the exception: chatbot framing, signposting, heading warm-ups, and passages that carry no content beyond their own rhetoric get cut, not rewritten, so some shrinkage from removing ceremony is expected. The test is "did any actual information disappear?", not paragraph count.
• Never invent specifics. Concrete details make text feel human, which makes fabricating them tempting; a made-up interviewee, study, or statistic is worse than any AI-ism. Use only specifics present in the source text or supplied by the user. When the text needs a concrete detail it does not have, ask for one or leave the sentence general.
• The final rewrite contains no em dashes or en dashes. They are among the most reliable tells; replace each with a period, comma, colon, parentheses, or a restructured sentence. Scan the final text for them before returning; any hit means it is not done.
• No emojis, no curly quotes, no mechanical boldface in the final rewrite.
• Match the register. For encyclopedic, technical, legal, or reference text, neutral and plain is the correct human voice; injecting opinions or first person there makes it worse, not warmer.

Core rules

Apply these principles first, then run the catalog to catch the specific tells they miss. Principles fix the prose; the catalog names the residue.

• Cut throat-clearing and filler. Openers that announce the point ("here's the thing", "it turns out", "the real question is") and phrases that add no meaning ("it is worth noting", "at the end of the day") delay the content; the text should start at its first real sentence. Why: the warm-up is where the reader's attention leaks, and its evenness is itself a tell.
• Break binary contrasts. "Not X, it's Y", "the answer isn't X, it's Y", "X isn't the problem, Y is" are telegraphed reversals; state Y directly and drop the negation. Why: the setup-reveal cadence is one of the most mechanical AI rhythms, and stating Y is shorter and truer.
• Prefer active voice with a named actor. Every claim should have a subject doing something; replace passive constructions and hidden actors with the person or thing responsible. Why: passive voice drains energy and lets the sentence dodge who actually acted.
• Choose specifics over vague declaratives. "The implications are significant", "the reasons are structural" announce importance without naming it; replace each with the concrete implication or reason, or cut it. Why: humans hoard particulars, models round them off, so a named specific reads human and a vague abstraction reads generated.
• Put the narrator in the room. "Nobody designed this", "people tend to" float above the scene; "you" and a concrete situation place the reader inside it. Why: the distant-lecturer voice is generic by construction, and proximity is what makes prose feel written by someone who was there.

Quick checks

Run this scan before the full rewrite; it surfaces the most common misses fast, then the catalog pass catches the rest.

• Does the text open by announcing itself ("here's what", "let's dive in") instead of starting with content?
• Any "not X, it's Y" or "the question isn't X, it's Y" contrast that could just state Y?
• Any sentence whose actor is hidden by passive voice or handed to an inanimate thing ("the data tells us", "the decision emerges")?
• Any vague declarative of importance ("the stakes are high", "this matters") with no specific named behind it?
• Any business jargon ("navigate", "deep dive", "circle back", "lean into") standing in for a plain verb?
• Any em dash, en dash, emoji, or curly quote (the hard-rule scan, also run again at the end)?

Curse-of-knowledge diagnostic

The checks above catch how the text is phrased; this one catches what the text wrongly assumes the reader already knows. It is a separate problem class: a passage can be free of every AI tell and still be unreadable because the writer skipped the context, the stakes, or the concrete example an outsider needs. Run it as its own pass; the fixes are supplying missing information, not rephrasing existing sentences, so do not fold it into the tells scan.

When a check fires, the fix obeys the hard rules: supply the missing term, stake, consequence, or example only from specifics in the source or from the user (never invent one), or flag the gap for the user to fill. Leaving the sentence general is better than fabricating the detail it lacks.

Workflow

1. Calibrate voice. If the user provides a writing sample (inline or a file path), read it first and note sentence-length habits, word register, punctuation tics, and how transitions are handled; the rewrite should replace AI patterns with that writer's patterns, not generic polish. Without a sample, default to a natural, varied voice fitting the content's register.
2. Quick-check pre-pass. Run the Quick checks scan above against the text; apply the Core rules to what it surfaces. Then run the curse-of-knowledge diagnostic as a separate pass, since it finds assumed-knowledge gaps the phrasing scan cannot. This catches the common misses before the slower catalog pass.
3. Load the catalog. Read references/ai-patterns.md and check the text against every pattern in it. The catalog is the skill; working from memory of it produces partial passes.
4. Judge before flagging. Apply the cluster rule below; confirm each flagged instance is a real tell and not legitimate prose.
5. Draft rewrite. Fix every confirmed instance. Read it aloud mentally: varied sentence length, simple copulas (is, are, has), specific over abstract.
6. Self-audit. Ask: what makes this draft still obviously AI-generated? List the remaining tells honestly; an even, mid-length cadence and too-tidy contrasts survive most first drafts.
7. Final rewrite. Address the audit findings, then run the hard-rule scans (dashes, emojis, curly quotes).
8. Deliver the final rewrite plus a short summary of the pattern groups removed. Show the intermediate draft and audit only when the user asks to see the work.

Detection judgment

The cluster rule. A single tell means nothing; clusters are the signal. One em dash is a punctuation choice; em dashes plus rule-of-three plus "vibrant tapestry" plus a generic upbeat conclusion is a confession. Weigh accordingly before rewriting a passage.

These are not reliable indicators on their own, so do not flatten them:

• Perfect grammar, consistent style, or correct complex formatting (professionals and templates exist)
• Formal or academic vocabulary in general; the catalog lists the specific overused words, and "ostensibly" is not one of them
• Common transition words in isolation; one "however" is not a tell, a pileup is
• Curly quotes alone (most editors auto-curl) or em dashes alone (journalists love them); both count only alongside other tells
• Bland prose without specific tells, mixed registers, or unsourced claims

Preserve signs of human writing. Over-editing destroys exactly what makes prose sound human. When you see these, lean toward leaving the passage alone:

• Specific, unusual, hard-to-fabricate detail; humans hoard specifics, models round them off
• Mixed feelings and unresolved tension instead of clean takes
• Era-bound slang, memes, or in-jokes; genuine asides, parentheticals, and self-corrections
• Varied sentence length, including fragments
• First-person choices the writer could defend

Voice and soul

Removing AI patterns is half the job; sterile, voiceless writing is just as obviously machine-made. When the content calls for a human voice (blog posts, essays, opinion, personal writing), check the draft for soullessness: every sentence the same length, no opinions, no uncertainty, no first person where it belongs, reads like a press release. The fixes:

• Have reactions, not just facts. "I genuinely don't know how to feel about this" is more human than a neutral pro/con list, when the author's voice supports it.
• Vary the rhythm. Short punchy sentences. Then longer ones that take their time getting where they are going.
• Let some mess in. Tangents, asides, and half-formed thoughts read as a person; perfect structure reads as an algorithm.

This section applies only where opinion belongs; see the register rule above.

Gotchas

• The self-audit step is load-bearing. Single-pass rewrites reliably keep the even cadence and tidy paragraphing that mark AI text even after the vocabulary is fixed; the "what is still obviously AI here?" question is what catches them.
• Voice matching beats voice improving. If the user's sample says "stuff" and "things", the rewrite says "stuff" and "things"; upgrading their vocabulary reintroduces the polish you were asked to remove.
• Hyphation has position rules. Keep compound hyphens before the noun (a high-quality report) and drop them after it (the report is high quality); uniform hyphenation everywhere is the machine pattern.
• Documentation should not narrate its own diff. Text that explains what the code used to do before a change (outside changelogs and migration guides) is a tell; describe the thing as it is.