/linkedin-humanizer

name: linkedin-humanizer
description: "Remove AI tells from LinkedIn posts/comments: 2026 vocabulary density, reveal bridges, staccato fragments, stacked triads, performed sincerity. Tiered rewriter plus `--mode audit` LinkedIn post auditor and `--mode profile` voice builder. Not for drafting from scratch (use linkedin-post-writer) or beating AI detectors. Keywords: humanize, de-AI, post audit, post auditor, audit before posting."

LinkedIn Humanizer V3

Rewrites any text to remove the AI tells that human readers notice and that LinkedIn's "AI slop" filter reacts to. Based on Wikipedia's "Signs of AI writing" taxonomy, the 2025-2026 stylometry literature, and our own length-controlled corpus. **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 (VUB IJEI 2026, Russell 2025), and light mechanical rewriting raises detectability (arXiv 2603.17522). No post-hoc edit reliably beats a Pangram-class detector, and detector scores on LinkedIn-length text (100-300 words) are noise. The real value is elsewhere: expert human readers cite vocabulary (53%) and sentence structure (36%) as what gives AI text away, and LinkedIn's July 2026 slop-report button costs a flagged post roughly 40% of its views. This skill removes what those readers and that filter react to.

What changed in V3

Evidence tier in brackets: [strong] = replicated across 2+ independent 2025-2026 studies or our own length-controlled corpus; [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: Geng & Trotta 2025]. 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 human rate [strong: Kobak Sci Adv 2025; Wu et al 2026; PNAS 2025]. AI vocabulary is also the one marker consistently reach-negative on LinkedIn in our own corpus (0.74-0.84 author-relative) [strong]. One marker in a paragraph is not a verdict. Three or more is.
• **Em dash is no longer a tell.** GPT-5.4 emits 1.43 per 1,000 words, below the 3.23 human baseline; 29% of human captions and 23% of top-creator LinkedIn posts in our corpus use one (author-relative ratio 1.09) [strong]. Zero em dashes is now its own tell (the writer is trying to look human). New rule: cap at about 1 per 100 words, replace excess with comma, colon, parentheses or a rewrite. Never a period.
• **Forced burstiness is the #1 2026 tell, not the fix.** LLM sentence-length variance is half of human [strong], but detectors do not score it, mechanical long/short alternation is a learnable humanizer fingerprint [weak: DAMAGE 2025], and on LinkedIn sentence-length variance is not an engagement lever in either direction (our corpus, n=397, within-creator: null to slightly negative) [strong]. "Short. Punchy. Done.", "No X. No Y. Just Z.", one-word paragraphs and "The result?" reveals are the current top tells. Pass 2 is now RHYTHM, not BREAK: fix machine-flat rhythm, never manufacture variance.
• **Rule of three is still a tell, at density.** Tricolon runs at 2x expert-human rate across 2026 frontier models [strong: arXiv 2604.19768]. Stacked, perfectly parallel triads and 3+ per post get scrubbed. One natural triple stays (26% of top human tweets have one).
• **Fingerprint injection was half wrong.** Named entities and concreteness are supported [strong: lower entity density in LLM text across 3 studies]; an odd-precision number with a referent in line 1 lifts likes 34% [vendor]. Bare numbers are not a discriminator, and inserted hedges and confessions backfire: performed hesitancy is 2x more common in LLM text than expert human text, and sincerity announcements ("let me be honest") are a named 2026 tell [strong: tropes.fyi false vulnerability; Schilke & Reimann 2025]. Pass 3 now asks for a flat, dated, uncomfortable fact instead.
• **Over-correction guard.** Humanizer output has its own fingerprint; "writing slightly worse on purpose" now reads as a tell [weak: DAMAGE 2025; slopotron]. 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.

See `sub-skills/rules-explainer.md` for per-rule justification, defenses, and citations, and `references/tier-rationale.md` §V3 for the evidence.

When to use

• Before publishing any AI-drafted post or comment (rewrite mode)
• Pre-publish review of a finished draft (audit mode, see `sub-skills/post-audit.md`)
• When a draft feels off and you can't pinpoint why

Input

Any text (post, comment, reply, DM). Optional: target voice samples (past human posts by the user).

Output

• Rewritten text with AI tells removed
• Diff showing what changed and why
• Per-paragraph tell density (markers per paragraph; 3+ triggered a rewrite)
• Reader-read confidence: "reads human", "mixed", "reads AI" (this is a reader-tell estimate, not a detector score)
• Tier applied (which mode was used)

Modes

# Default: forensic + strict (recommended for LinkedIn)
linkedin-humanizer <text>

# Forensic only: minimum-touch, just kill the leakage
linkedin-humanizer --mode forensic <text>

# Strict: forensic + density-scored 2026 vocabulary, reveal bridges, staccato (the LinkedIn-default config)
linkedin-humanizer --mode strict <text>

# Aesthetic: strict + style rules (single natural triads, passi