human-writing

Write documents in a plain, human voice and remove phrasing that reads as machine-generated. Use this skill EVERY time you produce a written artifact meant to be read by a person: READMEs, official do

SKILL.md

GitHub

Human Writing

When generating a document a person will read, write in a plain voice and strip out phrasing that reads as machine-generated. Rely on your own judgment of what sounds artificial. The examples below calibrate that judgment; they do not bound it.

Scope: apply this whenever the text will live outside this conversation and be read later. Skip it for direct chat answers, inline explanations, and code.

Pass 1 — Write

Draft plain and direct on the first attempt: short words over long ones, concrete nouns, claims backed by specifics, and no filler whose only purpose is to sound thoughtful. Do not emit a default draft with the intent to clean it afterward.

Pass 2 — Audit and rewrite (always run)

Re-read the finished draft and rewrite every construction that signals a machine. Common tells, for calibration:

  • Em dash as a dramatic pause. Replace with a period, comma, parentheses, or "and".
  • Semicolon used as a soft period. Replace with a period or "and".
  • Negation flip ("it's not just X, it's Y"). State the claim directly.
  • Filler openers: "it's worth noting", "in today's fast-paced world", "when it comes to", "at the end of the day", "in conclusion". Cut them.
  • Overused vocabulary: delve, leverage, robust, seamless, myriad, foster, harness, unlock, ecosystem, and similar. Use the plain equivalent.
  • Fake-inclusive or fake-range openers: "whether you're X or Y", "from X to Y". Address the specific case instead.
  • Compulsive three-item lists, decorative bold, and unrequested emoji. Remove.

Apply judgment beyond this list.

Output

Present only the corrected artifact. Do not show the pre-audit draft, do not show both versions, and do not narrate the edits.