Ig Humanizer

Instagram Humanizer V3

Rewrites any caption or carousel slide text to remove the AI tells that human readers notice, and audits a finished caption against the 2026 Instagram checklist. Based on Wikipedia's "Signs of AI writing" taxonomy, the 2025-2026 stylometry literature, our own Instagram caption corpus (n=284, prevalence only), and Instagram-specific patterns (the first-125 fold, the lowercase-casual caption register, sized hashtags, sends-and-saves structure). V3 (2026-09): recalibrated on 2026 evidence. Vocabulary is scored by density, em dashes are capped instead of banned (and the cap is generous here: 29% of human captions use one), 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 caption-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 on Instagram a caption that reads as a brand account earns neither the save nor the send. 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 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]. AI vocabulary appears in 10% of human captions in our corpus [strong], so one marker in a paragraph 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 [strong], and 29% of human Instagram captions contain one [strong: corpus]. A caption's em dash is never a tell on its own, and a blanket ban over-sterilises captions. New rule: cap at about 1 per 100 words (1-2 per caption), replace only the excess with a comma, colon, parentheses or a rewrite. Never a period (a split dash stacks fragments). Zero em dashes across a long caption is its own tell now.
• 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]. Captions are mid-length, so Pass 2 is an anti-uniformity guard only: fix a paragraph that reads machine-flat, never manufacture variance.
• Rule of three is still a tell, at density. Tricolon runs at 2x expert-human rate across 2026 frontier models [strong], and 23% of human captions contain one [strong: corpus]. So one natural triple with concrete items stays. Stacked, perfectly parallel or hollow triads and a third triad in a caption get scrubbed.
• Fingerprint injection was half wrong. Named entities and concreteness are supported [strong]; an odd-precision number with a referent in the first 125 chars is the strongest hook. 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", "real talk", "POV:" on something that is not a POV) 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 caption or carousel (rewrite mode)
• Pre-publish review of a finished caption (audit mode, see sub-skills/post-audit.md)
• When a caption feels off and you cannot pinpoint why

Input

Any text: a caption, the on-image text of a carousel, or a Reel caption. Optional: target voice samples (the user's past captions).

Output

• Rewritten text with AI tells removed
• A diff showing what changed and why
• Caption char count (flagging if the hook spills past the first 125 chars)
• Per-paragraph tell density (markers per paragraph; 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 Instagram-format issues
ig-humanizer <text>

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

# Audit - detection-only pass-fail review, no rewrite
# Runs the 2026 Instagram checklist: first-125 hook, caption length, sized
# hashtags, emoji limits, CTA quality, media reminder.
# Returns Blockers + Warnings + suggested fixes. See sub-skills/post-audit.md.
ig-humanizer --mode audit <text>

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

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 paragraph (or slide), not the word: count markers per paragraph, rewrite the paragraph 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 about 1 per 100 words).
• 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 third triad in a caption, phrase cleanups ("in today's fast-paced world", "game-changer", "level up", "dive in"), emoji storms, and dead closers ("what do you think?", "double tap if you agree").
• Instagram-format scrubs (always apply): first-125 hook that stands alone, sized hashtags, emoji limits, carousel slide-1 promise, caption length.

Pass 2 - RHYTHM (anti-uniformity guard only)

Detectors do not score burstiness, and captions are mid-length (our corpus median is well under 200 words), so rhythm is neither a reach lever nor a fix. What readers notice is the mechanical-uniformity tell (every line the same length, machine-flat) and, worse, the staged variance that second-generation humanizers add. So Pass 2 has two jobs: fix rhythm only where it reads machine-flat, and remove manufactured variance everywhere. It never adds variance as a tactic.

• Per paragraph: one genuinely long sentence next to a short one is fine and is what human variance looks like. Two or three mid-length sentences in a row are also fine. Edit only when every sentence in the paragraph runs the same length and reads flat, and then edit one sentence, not the paragraph.
• Standalone fragments: at most 2 per caption, total. "every time." once is a voice quirk. Three in a caption 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 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. One idea per line with blank lines between them is native caption formatting and stays. Fragment-for-drama inside those lines is the tell. Keep the layout, fix the sentences.
• Carousel slide text is short by design; never chop or pad a slide for rhythm. Never alternate long/short/long/short across slides or lines. That seesaw is the humanizer fingerprint.

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