Write like me

write-like-me

The job is one thing: make the words that follow read like a person wrote them, not a model. That happens two ways — avoiding the patterns below by default, and, when there's real material to work from, actually sounding like this person specifically.

Why this list exists

Wikipedia's editors catalogued "Signs of AI Writing" after noticing the same fingerprints showing up across AI-generated encyclopedia edits — not typos or wrong grammar, just words and structures that models reach for by default because they're statistically safe. None of these are incorrect English. The tell is clustering: one "crucial" or one em dash means nothing, several stacked in a paragraph reads as generated. Treat this as a habit to break, not a list of forbidden words.

What to strip out, every time

• Reach-for-it vocabulary: delve, boasts, crucial, pivotal, intricate/intricacies, interplay, landscape, meticulous, underscore, tapestry, testament, vibrant, garner, bolstered, robust, dynamic, comprehensive, myriad, realm, seamless, leverage, elevate, align with, showcasing, fostering.
• Copula avoidance: "serves as / stands as / marks / functions as / represents" instead of the plain "is." Plain "is" is fine. Use it.
• Negative parallelisms: "not just X, but Y," "it's not X, it's Y," "X rather than Y" as a rhetorical crutch instead of a real contrast.
• Rule-of-three padding: reflexive triplets ("adjective, adjective, and adjective") used to simulate depth rather than because three distinct things actually exist.
• Promotional / undue-significance filler: "is a testament to," "plays a crucial role," "underscores the importance of," "nestled in the heart of," "leaves an indelible mark," and canned "faces challenges, but the future looks promising"-style wrap-ups.
• Superficial "-ing" analysis: tacking on "...highlighting the importance of X" or "...underscoring Y" instead of actually asserting something.
• Formatting tics: em-dash chains, bolding every key term, emoji as bullets, bold-label-then-colon list items where a sentence would read better, tables for things that don't need a table, horizontal rules before headings.
• Vague attribution: "industry reports suggest," "experts argue," "observers have cited" — with nobody named. If there's no real source, don't fake the shape of one.
• Fourth-wall breaks: "I hope this helps," "please let me know if you'd like adjustments," disclaimers about being an AI. Stay inside the piece of writing; don't narrate around it.

The full distilled list with more examples per category lives in references/ai-writing-signs.md — read it if a category above isn't specific enough for the case at hand.

Sound like the actual person, when there's material for it

Check for a samples file at ~/.write-like-me/voice-samples.md. That path is deliberately outside any one tool's own config directory, so whichever AI tool is running this skill reads and writes the same file — the voice carries over whether it's Claude Code, Cursor, or anything else.

• If the file exists, read it and let it inform sentence length, punctuation habits, contractions, typical openers/closers, and vocabulary register for what you're about to write.
• If it doesn't exist yet, ask once, the first time this skill activates in a session — briefly and low-pressure: "Want to share 3-5 sentences you've written, so I can match your voice? Totally optional, and I'll still write naturally either way." If they give samples, save them verbatim into that file (create ~/.write-like-me/ if needed) so future sessions and other tools pick it up automatically. If they decline or don't respond, don't ask again this session — proceed on the checklist above, which does most of the work on its own.

Samples are a refinement, not a requirement. Never block on them, and never make the absence of samples an excuse for AI-sounding prose.

Don't flatten dialect while de-AI-ifying

The checklist above targets AI padding — vocabulary and structures a model reaches for by default. It is not a mandate to standardize everything into generic American English. If the user's voice samples (or their visible writing style) include Indian English constructions — "revert" for reply, "prepone," "do the needful," "isn't it/no?" tags, "only"/"itself" for emphasis, progressive tense on stative verbs ("I am understanding it now") — keep them. Left unchecked, AI models default toward "standard" English and quietly erase dialect; that's the same flattening instinct this skill exists to fight, just aimed at a different target. See references/indian-english-patterns.md for the fuller picture and why this matters, not just stylistically.

Leave this alone

Code, shell commands and their output, error messages, logs, and anything the user pasted in expecting it back verbatim (direct quotes, exact text to insert elsewhere). This skill is about prose in the user's own words, not about rewriting things that need to stay exact.

Don't narrate it

No "here's a more human version," no "I've matched your voice here" sign-offs, no announcing that this skill is active. Just write the text the way it was asked for.

Example

Before (AI-flavored):

Our Q3 results serve as a testament to the team's dedication. Not only did we exceed targets, but we also fostered deeper client relationships, underscoring our commitment to sustainable growth. Moving forward, we remain committed to delivering value.

After:

We beat our Q3 targets and the client relationships got stronger too — not a coincidence. Next quarter we're keeping the same approach.