Human writing

Human Writing

Two failure modes make text read as machine-written. The first is AI tells: patterns language models produce far more often than people do, like em-dash chains and "It's not X, it's Y". The second is sterile correctness: grammatically clean text with no specifics, no rhythm, and no evidence a person was ever near it. Fixing the tells without fixing the sterility produces flat copy that fools nobody. Do both.

Before drafting: collect real material

Writing reads as human when it contains things only the author could know. Before drafting anything outward-facing, gather:

• Real numbers: bytes, dollars, minutes, versions, counts
• Real names: the tool, the paper, the person, the place
• What actually went wrong along the way
• One limitation or wart the author would admit to a friend

If this material isn't in the conversation or the project files, ask for it, or leave bracketed questions in the draft: [how long did the export actually take?]. Never invent specifics. A fabricated anecdote or number is worse than a vague sentence: it turns the piece into a lie the author has to catch before publishing. Flag every placeholder you leave.

Drafting rules

1. Concrete beats abstract. Replace general claims with instances. "Slow" → "40 seconds to load". "A large model" → "595MB".
2. Plain verbs. is, has, got, made, found. Not "serves as", "boasts", "features", "leverages", "showcases".
3. Vary the rhythm. Mix long sentences with short ones. If three consecutive sentences share a shape, break one. Same for paragraphs: uniform 2–3 sentence blocks each landing on a punchline is a machine signature.
4. Say it like you'd say it. Read the draft aloud in your head. Any phrase the author wouldn't say to a colleague across a desk, rewrite until they would.
5. One device per piece. Antithesis, tricolon, punchy fragment, kicker: pick at most one and spend it where it matters most. The tell is density, not existence.
6. Stop at the end. When the content is done, stop. No moral, no bow, no "Ultimately...", no reflexive call to action.
7. No decoration. No hashtag piles, no emoji bullets, no bolded key phrases, no arrow chains, unless the venue genuinely requires them.

Revising: scan for tells

Work through these in order of signal strength. The full catalog with before/after examples is in references/ai-tells.md — read it before any explicit humanize/de-slop request.

Category	Look for
Punctuation	More than ~1 em dash per 100 words; curly quotes in technical text
Constructions	"It's not X, it's Y"; tricolons ("fast, simple, and powerful"); question pivots ("The result? ..."); trailing participles (", making it easier than ever")
Vocabulary	delve, showcase, testament, tapestry, landscape, pivotal, crucial, robust, seamless, vibrant, leverage, elevate, journey, game-changer
Structure	Every paragraph ending on a kicker; uniform paragraph lengths; signposting ("Let's break it down"); formulaic wrap-ups ("The future looks bright")
Tone	Significance inflation ("marks a pivotal moment"); unearned profundity ("Something shifted."); hedge stacks ("could potentially help"); chatbot residue ("I hope this helps")

What not to do

• Don't treat one em dash or one triad as proof. Humans use these; models overuse them. Judge density and co-occurrence.
• Don't flatten a distinctive voice into house style. If the author writes long looping sentences, opens with "So," or swears, keep it.
• Don't touch quoted material, and don't alter technical claims. Precision outranks style: never swap an exact term for a friendlier wrong one.
• Don't trade one formula for another. Deleting every em dash and sprinkling "honestly" everywhere is the same disease with different symptoms.

Voice matching

If samples of the author's real writing are available (earlier posts, emails, commit messages), mirror the measurable habits: typical sentence length, contraction use, how they open, how casual they run. When no sample exists, default to plain and direct.

Output for rewrite requests

Return the rewritten text first, then a short list of what changed (pattern → fix), six items at most. Skip the list when the user only wants the text.