| name | gtm-humanizer |
| description | Rewrite AI-generated text — especially DRAFTED EMAILS AND MESSAGES — so they read like a real person wrote them, not a chatbot. Use as the final pass on any email, Slack message, follow-up, case study, or customer-facing copy a GTM Superintelligence agent drafts, and whenever the user says "humanize this," "sounds like AI," "too robotic," "de-AI this," "sounds like ChatGPT," or "make it natural." Detects common AI-writing tells, scores how AI it sounds (0–100), and rewrites in a chosen voice profile. Bans em dashes by default. Auto-loads humanizer-context.md from the project root for brand/sender voice. |
| user-invocable | true |
| argument-hint | "your text" [--mode detect|rewrite|edit] [--voice casual|professional|technical|warm|blunt] [--file path] [--purpose email|message|marketing|general] [--score] |
| allowed-tools | ["Read","Write","Edit","Grep","Glob"] |
GTM Humanizer
Take text that reads like a chatbot produced it and rewrite it so a specific, busy human
clearly wrote it. Built for GTM: cold emails, follow-ups, Slack messages, handoff notes,
case studies — the copy that has to earn a reply, not just be grammatically fine.
The failure mode this fixes: text that's smooth, generic, over-helpful, and stake-free —
it could have been sent to anyone, by anyone, about anything. Human writing has a point
of view, specifics, and one clear ask.
How to run
- Load context. If
humanizer-context.md exists at the project root, read it first
and obey it — it defines the sender/brand voice and hard rules (in this repo it sets a
GTM sender voice and bans em dashes). Project rules win over defaults.
- Detect. Scan the text against
patterns.md (the catalog of AI-writing tells). List
each tell you find with the offending span.
- Score (0–100, lower = more human). Roughly: count distinct tells, weight the loud
ones (em dashes, "I hope this finds you well," hype adjectives, rule-of-three) higher.
Bands: 0–15 reads human · 16–35 mostly human · 36–60 mixed · 61–100 obviously AI.
- Rewrite in the chosen
--voice (see voice-profiles.md; default professional,
lean blunt for cold outreach). Fix every detected tell. Keep the author's real
meaning, facts, and any specific detail (names, numbers, quotes) — specificity is what
proves a human wrote it. Do not invent facts to sound human.
- Re-check the rewrite against the catalog. If new tells crept in, fix them. Report
the before/after score.
Modes
--mode detect — only report tells + score; don't rewrite.--mode rewrite (default) — return the humanized text.--mode edit --file <path> — rewrite in place in a file.
Hard rules (always, unless context overrides)
- No em dashes. Use periods or commas.
- No throat-clearing openers: "I hope this email finds you well," "I wanted to reach
out," "I'm excited to," "As an AI," "In today's fast-paced world."
- No hype adjectives: seamless, robust, leverage, synergy, game-changer, cutting-edge,
best-in-class, unlock, delve, elevate, tapestry, robustly.
- No rule-of-three padding ("streamline, optimize, and accelerate").
- One clear ask. Lead with the point. Cut anything that restates what you just said.
- Keep it the right length for the channel (email: short and skimmable; Slack: shorter,
no bold-text theatre; case study: quotes and outcomes over adjectives).
Output
By default return just the rewritten text (ready to paste/send). If --score or detect
mode, also print the tell list and the before→after score. Never add commentary the user
didn't ask for.
See patterns.md for the full tell catalog and voice-profiles.md for the voices.
