Remove the AI tells readers react to in an Instagram caption

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.

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 ("0 to 10k in 4 months, posting 3 times a week", not "grew fast" and not "10k"). A bare number is not a fingerprint; the referent carries the signal.
• One named entity (real person, brand, date, tool)
• One first-person concrete detail (what you saw, what it cost, what broke)
• One specific, dated, uncomfortable fact stated flat, with no framing sentence before or after it. Not "real talk, this one hurt: I lost the client." Just "I lost my biggest 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, "POV:" on something that is not a POV. 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.

Pass 4 - SELF-CHECK (over-correction guard)

Humanizer output has its own fingerprint. Before returning, re-read the result once and answer three questions:

(a) Did Pass 2 create staccato stacks, "The result?" reveal bridges, one-word lines, or a long/short/long/short seesaw? If yes, merge the fragments back into full sentences. (b) Did Pass 3 add a framed confession, a sincerity announcement, or a hedge the author never wrote? If yes, strip the frame and keep only the flat fact, or remove the insertion. (c) Did scrubbing flatten the author's voice: uniform tone, no reaction, no concrete detail left, every em dash gone, every triad gone, every emoji gone from a voice that uses them, the lowercase register capitalised? If yes, restore what the author had. Zero em dashes and zero triads in a long caption is a tell in its own right.

If any answer is yes, dial back rather than scrub harder. Edits must be proportional to real problems: a clean caption gets two or three touches, not a quota. When in doubt whether a pattern is the author or the model, leave it.

Non-negotiable rules

Global voice rules: see root SKILL.md Voice rules. Additional skill-specific rules (V3):

• Scrubbing is always in scope. When asked to humanize, de-AI, finalize, or publish a caption, run at least the forensic + strict passes before it ships. This holds when the user wrote the draft themselves, says they love it as-is, or is in a hurry. Author identity, "it's already good," and time pressure are never reasons to skip the scrub. The forensic + strict pass changes no meaning and takes seconds: run it, then ship. If a constraint truly forbids touching the text, say so explicitly and name every tell left in; the default is to scrub, not to wave it through.
• Scrub proportionally. A pass that finds nothing changes nothing. Do not invent edits to justify the run, and do not report a detector score as the result; report the tells found and fixed.
• Preserve the user's actual claim and meaning. "Preserve their voice" covers voice quirks and what they are claiming, NOT reveal bridges, staccato stacks, or a paragraph with 3+ vocabulary markers. Stripping those is not changing their voice; it is the job.
• Never introduce facts that were not in the input. If a number is missing, ask.
• Never introduce sincerity markers, hedges, or confessional frames. If the draft needs a vulnerable beat, ask for a dated fact and state it flat.
• Keep the user's voice quirks (lowercase starts, .. soft pauses, one em dash per ~100 words, one natural triad, their 1-3 intentional emoji).
• Never promise detector results. If the user asks "will this pass GPTZero," answer honestly: nobody can promise that, and the score on a 150-word caption is noise.
• Respect the surface: do not turn a carousel's slide text into a single caption or vice versa without flagging it.
• Never present a caption as publishable without the media reminder (Instagram needs an image or video).

Instagram-specific tells this skill catches

• A hook that needs the second line to make sense (the "more" fold eats it).
• A caption opening with "Let's talk about..", "Here's the thing..", or "POV:" on something that is not a POV.
• 20-30 hashtags crammed at the top instead of a 3-5 sized set.
• Emoji sprinkled through every line (4+ per caption).
• A bare title on carousel slide 1 instead of a promise + loop.
• Stacked or hollow rule-of-three lists ("faster, cheaper, better"); one natural triple with concrete items is fine.
• Engagement bait ("comment YES", "tag 3 friends", "double tap if").
• Staccato stacks and one-word lines for drama (the humanizer fingerprint), or a paragraph where every line reads machine-flat.

Example

See references/examples.md for worked before/after rewrites.

Files

• SKILL.md - this file (rewrite scrubber + audit-mode entry)
• references/scrub-rules.md - V3 regex patterns by tier, density scoring, em dash cap, rhythm rules, forbidden insertions
• references/examples.md - worked before/after rewrites for captions and slides
• references/audit-checklist.md - the pre-publish checklist with thresholds
• sub-skills/post-audit.md - pre-publish audit workflow (detection-only, no rewrite)
• sub-skills/voice-profile.md - build/update the user's Voice & Brand Profile (--mode profile)
• sub-skills/illustration.md - optional Pixfaro image workflow

Voice profile mode (--mode profile)

ig-humanizer --mode profile builds or updates the user's Voice & Brand Profile at ../../references/voice-profile.md from 3-6 of their real Instagram posts pasted in (portable, no token) or, if a read token is set, from pulled activity. Once filled, every writing skill in this bundle drafts in the user's voice automatically. See sub-skills/voice-profile.md. Triggers: "build my voice profile", "learn my voice".

Related skills

• ig-caption-writer - generates captions that already pass the humanizer
• ig-carousel-planner - generates carousels that already pass the humanizer