humanize-text

AI-written prose has a recognizable fingerprint. The loudest tell is the em dash (—), which humans rarely type but models emit constantly, followed by smart quotes and a handful of stock phrases. This skill removes those signatures while preserving the writer's meaning and voice. Apply it as the last pass over any prose you produce for the user to send or publish.

The goal is text that reads as if a competent person typed it in a normal editor, not text that has been scrubbed into blandness. Fix the tells; keep the substance.

1. Punctuation and Unicode (the mechanical, high-signal fixes)

These are the strongest AI markers. Replace them thoughtfully — choose the punctuation a human would actually have used, don't just delete.

• Em dash —: rewrite, don't blank out. Pick by role:
  • parenthetical aside → wrap in commas, or use parentheses
  • a pause before a payoff/explanation → colon
  • two independent clauses joined → split into two sentences, or use a semicolon
  • Example: We shipped it — finally. → We shipped it, finally.
  • Example: Three services — auth, billing, search — went down. → Three services (auth, billing, search) went down.
• En dash –:
  • number/date/page ranges → a hyphen, or the word "to" (2019–2023 → 2019 to 2023)
  • used as a dash → treat like an em dash above
• Smart/curly quotes “ ” ‘ ’ → straight " and '.
• Ellipsis … → three periods ..., or cut it (trailing ellipses read as AI hedging).
• Non-breaking spaces, zero-width chars, and other stray Unicode → normal ASCII space or delete.
• Leave ordinary hyphens in compound words alone (well-known, real-time, p99).

2. Phrasing tells (lighter touch — reduce, don't robotically purge)

Models overuse a stock vocabulary and a few sentence shapes. Trim them where they appear, but only if the replacement stays natural — don't mangle a fine sentence just to avoid a word.

• Inflated verbs/nouns: delve, leverage, utilize, foster, underscore, elevate, boast, harness, spearhead, tapestry, realm, landscape, testament, beacon. Prefer plain equivalents (use over utilize, show over underscore).
• Hollow intensifiers: seamless, robust, cutting-edge, game-changing, best-in-class, world-class, unlock, supercharge.
• Formulaic frames: "It's not just X, it's Y", "In today's fast-paced world", "At the end of the day", "When it comes to…" — cut or rephrase directly.
• Throat-clearing transitions stacked at sentence starts: Moreover, Furthermore, Additionally, In conclusion. Keep at most where they earn their place.
• Compulsive rule-of-three lists and perfectly parallel bullet phrasing — vary the rhythm; real writing is a little uneven.
• Emoji sprinkled into headings or bullets (unless the user's own style uses them).

3. What NOT to touch

Correctness matters more than de-AI-ing. Never alter:

• Code, code blocks, inline code, commands, file paths, URLs, regexes — a dash or quote there is syntax, not style.
• Verbatim quotations, names, citations, and literal strings the user must keep exact.
• Technical notation: ranges/signs where a hyphen is meaningful (-1, x-y), version numbers, IDs.
• The meaning, claims, or facts of the text. This is a style pass only.

Applying it

• When you generate prose in a normal response, apply these fixes inline before you present it — the user should just receive clean text.
• When editing a file of prose, make the edits directly.
• For a large plain-prose file, scripts/humanize.py <file> does the safe mechanical pass (dashes, quotes, ellipses, stray Unicode) and skips fenced code blocks; the phrasing work in section 2 still needs your judgment.
• If a change would distort meaning or the user explicitly wants a dash/quote kept, leave it and say why.