X Humanizer

X Humanizer V3

Rewrites any tweet or thread to remove the AI tells that human readers notice, and audits a finished draft against the 2026 X ranking checklist. Based on Wikipedia's "Signs of AI writing" taxonomy, the 2025-2026 stylometry literature, and our own length-controlled X corpus (n=445). 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 tweet-length text (under 300 words) detector scores are noise. The real value is elsewhere: expert human readers cite vocabulary (53%) and sentence structure (36%) as what gives AI text away, and X readers punish it with the ratio, the quote-dunk, and the scroll. This skill removes what those readers 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]. 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]. In our X corpus AI vocabulary appears in 14% of top tweets and those tweets earn 0.58x the median engagement [strong]. One marker in a tweet is not a verdict. Three is.
• Em dash is no longer a tell; the density is. GPT-5.4 emits 1.43 per 1,000 words, below the 3.23 human baseline [strong]. On X specifically em dashes are rare in top tweets (11%) and those tweets earn 0.52x the median [strong: corpus], so the cap here is tight: at most one per tweet, and none in a tweet that does not need one. Replace the excess with a comma, a colon, .., or a rewrite. Never a period (a split dash stacks fragments).
• Forced burstiness is the #1 2026 tell, not the fix. Mechanical long/short alternation is a learnable humanizer fingerprint [weak], and on X the rhythm rule flips with length: uniform rhythm wins on short posts (about 75 words, 1.7x median engagement for low-variance tweets) and natural variance only helps on long threads (about 430 words, 1.8x) [strong: corpus, length-controlled]. So Pass 2 never forces variance on a single tweet, and on a thread it only removes manufactured variance and un-flattens what reads machine-flat. "Short. Punchy. Done.", "No X. No Y. Just Z.", one-word tweets for drama and "The result?" reveals are the current top tells.
• Rule of three is still a tell, at density. Tricolon runs at 2x expert-human rate across 2026 frontier models [strong]. 26% of top human tweets contain exactly one [strong: corpus], so one natural triple with concrete items stays. Stacked, perfectly parallel triads and a second triad in the same tweet get scrubbed.
• Fingerprint injection was half wrong. Named entities and concreteness are supported [strong]; an odd-precision number with a referent in line 1 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 ("let me be honest", "unpopular opinion:" on a popular take) 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 publishing any AI-drafted tweet or thread (rewrite mode)
• Pre-publish review of a finished draft (audit mode, see sub-skills/post-audit.md)
• When a draft feels off and you cannot pinpoint why

Input

Any text: a single tweet, a thread (with or without --- breaks), a reply, or a quote-tweet draft. Optional: target voice samples (the user's past tweets).

Output

• Rewritten text with AI tells removed
• A diff showing what changed and why
• Per-tweet char count (flagging anything over 280, emoji counted as 2)
• Per-tweet tell density (markers per tweet; 3+ triggered a rewrite)
• Reader-read confidence: "reads human", "mixed", "reads AI" (a reader-tell estimate, not a detector score)

Modes

# Default: scrub AI tells (forensic + strict) and fix X-format issues
x-humanizer <text>

# Forensic only - minimum touch, just kill model leakage
x-humanizer --mode forensic <text>

# Audit - detection-only pass-fail review, no rewrite
# Runs the 2026 X checklist: 280-char fit, first-line hook, hashtag/emoji
# limits, link placement, thread tap-through, goal clarity.
# Returns Blockers + Warnings + suggested fixes. See sub-skills/post-audit.md.
x-humanizer --mode audit <text>

# Profile - build/update the user's Voice & Brand Profile. See the section below.
x-humanizer --mode profile

# Interview - fill the Story Bank: what you have to say, not how you say it.
# 20-40 minutes, resumable, or --mode interview post for one focused topic.
x-humanizer --mode interview

The four passes

Pass 1 - SCRUB (score, then delete or replace)

Apply the tiered catalogs in references/scrub-rules.md. The unit of judgement is the tweet, not the word (in a Premium long post over 280 chars, the paragraph): count markers per unit, rewrite the unit at 3+, leave a single marker alone unless it is a reveal bridge, negative parallelism, a sincerity marker, or forensic leakage.

• Forensic (always on): real model leakage no human types. AI tool markers (oaicite, contentReference, turn0search0), knowledge-cutoff disclaimers ("As of my last update"), template blanks ([Your Name]), and em dashes above the cap (more than one in a tweet).
• Strict (default on): what readers react to. The durable 2026 vocabulary set scored by density (significant, crucial, notably, particularly, comprehensive, insights, robust, leverage, foster, landscape, nuanced, streamline, elevate, empower), grammar markers (nominalisations, sentence-opening "-ing" clauses), the 2026 model-idiom layer (quietly, "X matters.", compound, "a signal", "the work", "built different", "let that sink in"), reveal bridges on a single hit ("The result?", "Here's what", "Stop X, start Y", "plot twist:"), all forms of negative parallelism, stacked or perfectly parallel triads and any second triad in a tweet, phrase cleanups ("in today's fast-paced world", "game-changer", "deep dive"), and dead closers ("what do you think?").
• X-format scrubs (always apply): 280-char fit with emoji as 2, hashtag and emoji limits, link placement, first line that stands alone.

Pass 2 - RHYTHM (never force it)

Detectors do not score burstiness. On X the corpus says rhythm depends on length: uniform rhythm wins on short posts and natural variance only helps on long threads. So Pass 2 has two jobs: remove manufactured variance everywhere, and un-flatten only a long thread that reads machine-flat. It never adds variance as a tactic.

• Single tweet, reply, or quote tweet: do not touch the rhythm. A tweet of three same-length sentences is how top tweets read (1.7x median engagement for uniform rhythm at about 75 words). Never insert a fragment, never chop a sentence to "add punch".
• Threads: a mix of tweet lengths that arises from the material is fine and is what human variance looks like. Edit only when every tweet runs the same length and reads flat, and then let the tweet carrying the most content take one real clause, once. Never insert a 3-word "punch tweet" for rhythm; the inserted punch is the humanizer fingerprint.
• Standalone fragments: at most 1 per tweet and 2 per thread. "Every time." once is a voice quirk; three in a thread is a pattern.
• Banned outright (rewrite as full sentences): "The X? Y." reveals; "No X. No Y. Just Z."; "All the X. None of the Y."; "Simple. Effective. Easy." adjective stacks; one-word tweets or lines for drama ("Still." "Exactly."); pseudo-Socratic Q&A ("Why? Because..."); "Short. Punchy. Done." staccato runs. Fragment runs are the tell.
• Layout is not rhythm. A hard return between two short lines is native X pacing and stays. Fragment-for-drama inside those lines is the tell.
• Never alternate long/short/long/short across a thread. That seesaw is the humanizer fingerprint.

The check is "did I add a staccato pattern, and does any long thread read machine-flat", not a variance number.

Pass 3 - ADD (human fingerprints)

Require where the content allows:

• One odd-precision number WITH a named referent: who, what, when, or what it cost ("$4,730 in Vercel overages, March invoice", not "$5k" and not "massive costs"). A bare number is not a fingerprint; the referent carries the signal.
• One named entity (real person, company, date, tool)
• One first-person concrete detail
• One specific, dated, uncomfortable fact stated flat, with no framing sentence before or after it. Not "not gonna lie, this one hurt: we lost the client." Just "We lost Carta as a client on 14 Feb." The fact carries the vulnerability. The frame turns it into performed sincerity, which readers now read as the tell.
• The lowercase-casual register if the voice calls for it

Forbidden as openers or pivots (sincerity announcements, a named 2026 tell): "let me be honest", "I'll be real", "honestly?", "to be direct", "the honest version is", "real talk", "not gonna lie", "ngl", "can I be vulnerable for a second", "unpopular opinion:" as a preface to a popular one. Also forbidden as insertions: hedges the author did not write ("perhaps", "I might be wrong but", "it seems"). Performed hesitancy is 2x more common in LLM text than in expert human text; adding it makes the draft read more AI, not less.

If the input lacks these, ask the user for a number, name, or moment. Do not fabricate.