Humanizer

Humanizer: Remove AI Patterns + Inject Human Quality

Full routing notes

Removes AI writing patterns and injects human quality — specificity, burstiness, and precise emotional language. Use when editing drafted content, cleaning up LinkedIn posts, product copy, or blog content so it doesn't sound like AI. Two-phase process: Phase 1 removes 33 AI tells (significance inflation, AI vocabulary, em dash overuse, contrastive negation, vague attributions, soulless structure); Phase 2 injects humanity (sentence length variation, specificity via "what kind?" technique, precise emotional language). An optional personal-brand mode (below) adapts the two phases to a specific author's voice, using a scoreable voice-schema gate and a deterministic eval harness. Do not use for generating new content from scratch, fact-checking, SEO scoring, or research. Does not use humanizer tools (gpthuman.ai, WriteHuman) — those are detectable; this skill does structural rewriting.

Two-phase editing skill. Phase 1 removes AI tells. Phase 2 injects specificity, burstiness, and emotional precision. Clean-but-generic is still a failure — both phases are required.

Setup

Load both reference files before starting:

1. references/patterns-reference.md — 33 AI tell patterns with before/after examples
2. references/specificity-techniques.md — "What kind?" technique, 7 levers, Emotions Wheel

Both are mandatory. Do not skip either.

Brand Context Intake

Before editing, identify the context. If not supplied, ask:

"Which context is this for? (a) Personal brand — LinkedIn/X/blog, (b) Analytical deliverable (report/memo/brief), (c) Generic content"

Personal-brand mode (example — adapt to your own voice)

When the draft is the author's own LinkedIn or X post, this mode wraps the two generic phases with a specific voice and a scoreable gate. It only rewrites an existing draft; generating a post from scratch is out of scope for this skill.

Required inputs: draft and channel (linkedin | x). Ask for either if missing. Optional: purpose, known_context (facts/examples that must be preserved).

Load, in order, before editing:

1. references/voice-schema.md — a scoreable quality gate template. Fill in the "Banned Vocabulary" / "Forbidden Phrases" / dimension descriptions with the author's own rules, or use it as shipped as a reasonable generic default.
2. Any voice-reference notes the author already maintains (brand guidelines, prior approved posts, a style doc). If none exist yet, ask 2-3 questions about tone, audience, and what they never say, and note the answers for next time.

Workflow:

1. Preserve meaning — capture the core claim, audience, concrete facts, personal details, caveats. Mark facts that must not change. If the draft lacks enough factual/personal material to feel author-specific, say so under Residual risks rather than inventing details.
2. Substance & structure — lead with the point, cut padded intros / fake balance / list symmetry that does not serve the argument. LinkedIn keeps a considered prose rhythm unless clearly tactical; X compresses to one idea per tweet.
3. Voice & humanize — run Phase 1 + Phase 2 below, then apply the author's voice rules from step 1: precise, direct, evidence-first; confident but not certain; no corporate-polish theater, no generic AI commentary, no fake certainty; no em-dashes if that's one of their rules.
4. Score & gate — score the output against references/voice-schema.md. When editing files/fixtures, run the deterministic harness:

   python scripts/eval_voice.py fixtures --format markdown
   

   If not run, state that the scorecard is manual-only.
5. Capture feedback — on approve/reject/manual-edit, save before/after to fixtures/before-after/ (approved), fixtures/aiish/ (rejected), or fixtures/gold/ (strong published). This builds a calibration set over time — don't rely on chat history alone.

Output: Revised Draft (2 variations) · Scorecard (vs schema) · Deterministic-check summary · Change notes · Residual risks · Publishability verdict. Never publish or send on the author's behalf — drafts are review inputs for a human.

Phase 1: Remove AI Tells

Load references/patterns-reference.md. Scan for all 33 patterns. Rewrite each flagged section.

Priority flags (most common, fix first):

• Em dashes (—): Do a mechanical character search for U+2014 before anything else. Every instance outside a code block must be replaced. This is not optional and not stylistic — em dashes are one of the strongest AI detection signals and must be eliminated completely from prose. Replace with: comma (asides), period (strong breaks), colon (heading label-separators). Zero exceptions outside code blocks.
• AI tell-words: delve, unpack, leverage, tapestry, landscape, pivotal, comprehensive, robust, seamless, furthermore, moreover, genuine, genuinely, "it's worth noting," "this underscores"
• 2025-26 cluster (the delve-era words are fading; these replaced them): unlock, elevate, holistic, nuanced, resonate, align, transformative, dynamic, empower, streamline, harness, paramount, meticulous
• Contrastive negation ("It's not just X, it's Y" / "not merely... but...") — the strongest 2025-26 tell. Delete the negated clause, state the claim directly.
• Self-answered rhetorical questions ("Why does this matter? Because...") — merge into one declarative
• Theatrical transitions: "Here's the thing:", "The result?", "Let's break it down:", "But here's the kicker:", "Ultimately,"
• "Aims to" / "seeks to" openers ("This article aims to explore...") — just start with the claim
• Fake citations: "research shows," "experts say," "studies suggest" without a named source
• Significance inflation: "marks a pivotal moment," "enduring testament," "evolving landscape"
• Copula avoidance: "serves as," "stands as," "functions as" → replace with "is"/"are"
• Sycophantic artifacts: "Great question!", "I hope this helps!", "Let me know if..."

Phase 2: Inject Human Quality

Load references/specificity-techniques.md. Run all four injection passes:

Pass 1 — Burstiness Find 3+ consecutive sentences of similar length. Break one up or combine two. Target: mix of short punchy sentences (6–8 words) and longer ones. Low burstiness is the strongest AI detection signal.

Pass 2 — Specificity injection For every key claim: ask "what kind?" three times. If the noun or verb can't answer "what kind?", it's not earning its place. Apply the 7 specificity levers (see reference file). Move the content from V1–V2 specificity to V3–V4.

Pass 3 — Emotional precision Find every emotional or evaluative word (good, bad, frustrated, excited, important, effective). Apply the Emotions Wheel: replace with the precise emotion. "I felt bad" → "I felt embarrassed" or "I felt resentful" — each tells a different story.

Pass 4 — Read-aloud test Read the output aloud. Anywhere you stumble or the words feel unnatural in your mouth — fix those. This catches what pattern-matching misses.

Output Format

1. The rewritten text
2. Brief change summary: which patterns removed (Phase 1) + which injections made (Phase 2)

For personal brand context: deliver 2 variations unless the user explicitly asks for one.

Rules

• Load both reference files — mandatory, not optional
• Em dash search is the first action in Phase 1. Search for — (U+2014) before reading for other patterns. Replace every instance outside a code block. Do not proceed until the count is zero.
• Never deliver Phase 1 only — clean-but-generic is still a failure
• Never do surface-level word substitution — fix the underlying claim, not just the word
• Never add chatbot artifacts to output ("Here is the revised text", "I hope this helps")
• Ask for brand context if not supplied — quality gates differ
• Personal brand: 2 variations by default
• When a noun/verb can't answer "what kind?", flag it explicitly — don't silently substitute
• Humanizer tools (gpthuman.ai etc.) are now detectable — never recommend them

Learning Loop (every run, automatic — do not ask first)

Run at the end of EVERY invocation. These are low-risk reversible writes; do them, then mention them in one line of the output.

• If the user approves the output — save the before/after pair to fixtures/before-after/ (personal-brand mode) or as an example in references/patterns-reference.md under "Saved Examples" (generic mode)
• If the user rejects the output as still AI-ish — save it to fixtures/aiish/ and append a dated entry to references/learnings.md (What Has Failed) naming which pattern survived the pass
• If the user identifies a new pattern — add it to references/patterns-reference.md as the next numbered pattern, same turn
• If the user says "never do X again" — add it to the Rules section above and note it in the relevant reference file
• If the user corrects a specific mistake — update the relevant pattern's description to prevent recurrence, and log the correction in references/learnings.md
• If the run surfaced nothing (clean pass, no user reaction yet) — skip silently; no empty entries

Credits

Phase 1 pattern taxonomy originally based on Wikipedia's "Signs of AI writing" guide via github.com/blader/humanizer (MIT licensed). This version extends it with Phase 2 (specificity/burstiness/emotional-precision injection), a personal-brand mode, and a deterministic eval harness.