/tt-humanizer

name: tt-humanizer
description: "Remove the AI-script tells viewers hear in a TikTok spoken script and caption: 2026 vocabulary by density, reveal bridges, staccato stacks, stacked triads, performed sincerity, written-not-spoken phrasing, \"hey guys\" filler; caps em dashes. Includes --mode audit pre-film check (hook, completion design, caption fit) and --mode profile. Not for beating AI detectors (no edit reliably does). Not for writing from scratch (use tt-hook-scripter). Keywords: humanize script, de-AI, audit before filming."

TikTok Humanizer V3

Rewrites a spoken script (and caption) to remove the AI tells that viewers hear, and audits a finished draft against the 2026 TikTok checklist before you film. The problem this solves is specific to video: a script that reads fine on the page can sound robotic out loud. Written-not-spoken phrasing, perfect parallelism, and AI vocabulary all expose themselves the second a human says them to camera.

Based on Wikipedia's "Signs of AI writing" taxonomy, the 2025-2026 stylometry literature, our own short-form corpora (X, Threads, Instagram captions), and TikTok-specific spoken patterns (the muted-first hook, the no-intro open, completion-rate structure). **V3 (2026-09):** recalibrated on 2026 evidence. Vocabulary is scored by density, em dashes are capped instead of banned, forced rhythm is now a tell instead of a fix, and there is an over-correction guard.

**What this skill does not do:** it does not make text "pass" GPTZero, Pangram, Turnitin or Originality. Those are trained classifiers keyed on the instruction-tuning style signature; prompt-style "sound like a real person" rewrites are caught 92-95% of the time, and light mechanical rewriting raises detectability. On script-length text (under 300 words) detector scores are noise, and nobody runs a detector on a video anyway. The real value is elsewhere: expert human readers cite vocabulary (53%) and sentence structure (36%) as what gives AI text away, and on TikTok a script that sounds read loses the viewer inside the first 3 seconds. This skill removes what those viewers react to.

What changed in V3

Evidence tier in brackets: [strong] = replicated across 2+ independent 2025-2026 studies or our own corpora; [vendor] = single platform or vendor dataset; [weak] = one study or expert-panel report.

• **Vocabulary moved from a delete-list to density scoring.** The 2023-24 words

(delve, tapestry, realm, journey) are decaying as humans avoid them [strong]. The durable 2026 markers are common words (significant, crucial, notably, comprehensive, insights, robust, leverage, foster, landscape, nuanced, streamline, elevate) plus grammar: nominalisations and "-ing" clause openers at 5.3x the human rate [strong]. Spoken, they are worse: nobody says "leveraging" to a camera. One marker in a script beat is not a verdict. Three is.

• **Em dash is no longer a tell.** GPT-5.4 emits 1.43 per 1,000 words, below

the 3.23 human baseline, and 29% of human captions on sibling platforms use one [strong]. In a spoken script a dash is only a breath mark the speaker sees, so it is never a tell there (`..` reads better on a teleprompter). In the caption: cap at about 1 per 100 words. On an on-screen card (3-7 words): at most one, and a card rarely needs one. Replace the excess with a comma, colon, `..` or a line break. Never a period.

• **Forced burstiness is the #1 2026 tell, not the fix.** Mechanical

long/short alternation is a learnable humanizer fingerprint [weak], and "Short. Punchy. Done.", "No X. No Y. Just Z.", one-word lines for drama and "The result?" reveals are the current top reader-cited tells [strong]. Spoken lines are naturally short, so Pass 2 is an anti-uniformity guard only: it makes the script sayable (contractions, one breath per line) and fixes a teleprompter-flat run, but it never inserts a punch line for rhythm.

• **Rule of three is still a tell, at density.** Tricolon runs at 2x

expert-human rate across 2026 frontier models [strong], and a perfect tricolon read aloud ("learn, grow, succeed") is the most audible tell there is. Stacked, perfectly parallel or hollow triads get scrubbed. One natural triple with concrete items stays (22-26% of top human posts have one).

• **Fingerprint injection was half wrong.** Named entities and concreteness are

supported [strong]; an odd-precision number with a referent in the hook is the strongest opener. Bare numbers are not a discriminator, and inserted hedges and confessions backfire: performed hesitancy is 2x more common in LLM text, and sincerity announcements ("not gonna lie", "let me be honest", "storytime" with no story) are a named 2026 tell [strong]. Pass 3 asks for a flat, dated, uncomfortable fact instead.

• **Over-correction guard.** Humanizer output has its own fingerprint [weak].

Pass 4 checks whether Passes 1-3 introduced the very patterns they were meant to remove. Edits are proportional to real problems. When in doubt, leave it.

When to use

• Before filming any AI-drafted spoken script (rewrite mode)
• Pre-film review of a finished script + caption (audit mode, see

`sub-skills/post-audit.md`)

• When a script "reads fine but sounds off" when you say it out loud

Input

A spoken script (the hook line plus the body), optionally the caption, and optionally voice samples (the user's past scripts or how they actually talk).

Output

• Rewritten script that sounds spoken, not written
• A diff showing what changed and why
• Caption char count (flagging over 2,200) when a caption is included
• Per-beat tell density (markers per script beat or caption paragraph; 3+

triggered a rewrite)

• Reader-read confidence: "sounds human", "mixed", "sounds read" (a

viewer-tell estimate, not a detector score)

Modes

# Default: scrub AI tells (forensic + strict) and fix spoken-word issues
tt-humanizer <script>

# Forensic only - minimum touch, just kill model l