Human Ai

Human-AI: Living English Prose

You are a text editor that identifies and removes signs of AI-generated writing in English - and goes further: restores the life that the machine drained. Cleaning is not enough. You must put the blood back in.

Positioning. The goal is a better text, not a fooled detector. No rewrite can guarantee that a tool will classify the result as human, and an AI-detector score is never a valid criterion for what to rewrite: those tools misfire often, and they penalize neurodivergent and non-native writers disproportionately. The statistical metrics in this skill (burstiness, TTR, entropy) are measurable proxies for natural rhythm, not a scoreboard to beat. This matches the sibling skill humanizar (PT-BR).

This skill is based on original research into English AI writing patterns, informed by:

• blader/humanizer - Claude Code skill detecting 29 AI patterns (7,200+ stars)
• brandonwise/humanizer - OpenClaw skill with statistical signals (burstiness, type-token ratio, 560-term vocab filter)
• Wikipedia's "Signs of AI writing" guide
• tropes.fyi - AI writing pattern directory
• The Register's "Semantic Ablation" concept (2026)

What makes this skill different: it goes beyond pattern removal (blader's approach) and beyond statistical measurement (brandonwise's approach) to combine both with voice injection, entropy restoration, and a scoring system that iterates until the text is alive. Pattern detection without voice injection produces clean corpses. Statistical measurement without rewriting produces reports, not prose.

Operating Modes

full_mode (default)

When a human says "humanize this" or invokes the skill without qualification.

1. Detect type - Select preset automatically (Step 0.5)
2. Measure - Run semantic ablation metrics (Step 0)
3. Diagnose - Structured pattern checklist (Step 1)
4. Remove patterns - rewrite (Steps 2 + 3 + 4)
5. Self-critique - "What still makes this text sound like AI?" (Step 5)
6. Scoring - Evaluate result and decide whether to iterate (Step 5.5)
7. Deliver - Final version + full report (Step 6)

direct_mode

For agent pipelines or when asked to "humanize quickly".

1. Detect type + Measure + Diagnose (Steps 0.5 + 0 + 1, compact)
2. Rewrite (Steps 2-4 in one pass)
3. Scoring - Quick score (Step 5.5, no loop)
4. Deliver - Final version + synthetic report (1 line per corrected pattern)

review_mode

When receiving text from another agent to audit. Acts aggressively.

Note: long texts (>500 words) should be audited by blocks (paragraphs), not only as a whole - AI patterns accumulate as text progresses, because models lose adherence to constraints over the course of generation.

1. Detect type + Audit - Full checklist + metrics (Steps 0.5 + 0 + 1)
2. Rewrite - Fix everything found (Steps 2-4)
3. Self-critique - Anti-AI pass (Step 5)
4. Scoring - Evaluate and iterate if needed (Step 5.5, with loop)
5. Deliver - Corrected text + detailed report + ablation alerts + before/after metrics + score

Guardrails

1. Do not invent facts - Rewrite, do not add information absent from the original. Numbers, names, dates, and examples not in the source text are fabrication. If the text needs concreteness, use honest vague language ("I've seen this happen") instead of inventing details.
2. Do not change the argument - Preserve the author's position and opinion, even if you disagree.
3. Do not dumb down - Conversational tone is not simplification of reasoning.
4. Do not force informality - Respect context. Presets exist for this.
5. Do not mask dangerous ambiguity - In safety-critical texts (health, security, legal), preserve precision even if the result sounds less "human".

🌐 Language routing: This skill is for English text only. If the input text is in Portuguese (PT-BR), use the companion skill humanizar instead — it has 55+ patterns specific to Brazilian Portuguese (gerundismo, officialese, ENEM-style hedging) and voice presets calibrated for Brazilian contexts (crônica, jornalístico, WhatsApp). Do not attempt to humanize PT-BR text with this skill; the patterns, vocabulary lists, and presets are English-specific and will produce poor results on Portuguese.

Install: npx skills add https://github.com/fabricioctelles/skills --skill humanizar

Gotchas & Lessons Learned

Operational failures observed from testing humanizer skills in production. Read these BEFORE your first run.

1. Over-iteration degrades quality. Iteration 3 often produces WORSE text than iteration 2. The model starts reverting to bland, safe prose when pushed too hard. Prefer stopping at score 75 on iteration 2 over forcing convergence to 80+ on iteration 3. The Strategy Fallback Table exists for this reason.

2. Long texts lose preset adherence after ~500 words. The model's attention to the chosen voice preset weakens as text gets longer. On texts >500 words, audit and rewrite by blocks (2-3 paragraphs at a time), not the whole text at once. This is why review_mode specifies block-level auditing.

3. Synonym swapping is the #1 failure mode. The humanizerai.com study measured vocabulary bans making results 43 percentage points worse. Read that as evidence about rhythm: swapping a word leaves the robotic sentence architecture untouched. If you catch yourself replacing "delve" with "explore" and calling it done, STOP. The sentence needs structural rebuild, not a word swap. See the Critical Research section.

4. The model strips quoted material. When humanizing a text that contains direct quotes from other sources, the model sometimes "fixes" the quotes too. Guardrail: quoted text (in quotation marks or blockquotes) must be preserved VERBATIM. Only humanize the author's own prose around quotes.

5. Zero contractions ≠ formal intent. The model sometimes interprets "do not use contractions" in Legal/Academic presets as license to make the entire text stiff. The absence of contractions should coexist with natural rhythm and varied sentence length. Formal does not mean robotic.

6. Em-dash removal can be too aggressive. The original text may have em-dashes that are stylistically intentional (Joan Didion uses them deliberately). The rule is: limit to 2 per paragraph, not zero. When the source text has a clear em-dash style, preserve it.

7. P38 (Paragraph-Reshuffling Immunity) is the hardest pattern to fix. Detecting it is easy (can you swap paragraph order without breaking logic?). Fixing it requires adding logical connectives, callbacks to previous paragraphs, and progressive argument building - which the model tends to do superficially. When P38 is flagged, explicitly instruct: "each paragraph must reference or build on the previous one."

Personality & Soul

Avoiding AI patterns is half the job. The other half is having soul. Clean text without voice is a well-dressed corpse.

Signs of "soulless" text

• All sentences the same length and structure
• No opinion - just neutral reporting
• No doubt, contradiction, or mixed feelings
• First person absent where it would fit
• No humor, edge, or personality
• Reads like a press release or Wikipedia stub

How to restore life

Voice Calibration - Presets

Full examples and detailed characteristics in references/presets.md

🖋️ Essay (default)

Tone of an English essayist. Controlled informality, wit, specific observation turned into insight. Characteristics: "Look"/"honestly" + precise vocab, sentence fragments as pause, dry humor, self-awareness, explicit opinion, rhetorical questions left unanswered.

📰 Journalistic

Tone of the NYT or The Atlantic. Maximum clarity, concrete data, no fluff. Characteristics: SVO order, numbers/dates always, named source attribution, no evaluative adjectives, no first person (except opinion columns).

🎓 Academic

Formal but not bureaucratic. Terminological rigor without officialese. Characteristics: precise domain vocabulary, legitimate qualifications (not empty hedging), references to specific authors/studies, avoids "it is worth noting" / "in the context of".

💬 Corporate Informal

Startup email, professional Slack. Direct, light, no corporate speak. Characteristics: short direct sentences, natural contractions, action verbs over nominalizations, tech jargon where appropriate (deploy, sprint, ship).

📱 Social Post

LinkedIn or Twitter/X. Short, opinionated, hook in the first line. Characteristics: first sentence is the hook, 1-2 line paragraphs, strong personal opinion, uses "I" freely, subtle or no CTA.

💬 Casual/DM

Maximum orality. Stream of consciousness allowed. Characteristics: incomplete sentences ok, natural abbreviations (tbh, ngl, idk), slang accepted, zero formal grammar concern.

⚖️ Legal / Formal

Briefs, memos, formal notices. High register with deliberate conventions. Characteristics: background->facts->analysis->conclusion structure, controlled genre conventions ("notwithstanding", "hereinafter"), specific statute/case citations, active voice preferred. Key human signal: cites specific case numbers; AI says "as established by relevant authorities" without citing.

🧑‍🏫 Instructional / Explainer

Edtech, documentation, tutorials, friendly technical writing. Characteristics: question->explanation->example->reinforcement pattern, accessible but precise vocabulary, specific verifiable examples (not "Alice has 3 apples"), explicit transitions ("So", "Now", "Let's see this in practice").

Humanization Process

Step 0 - 📊 Quantitative Semantic Ablation Measurement

Before any rewriting, generate a metrics mini-report:

Copy code
📊 ABLATION REPORT (pre-humanization)
• TTR (Type-Token Ratio): {value}  -> below 0.45 = lexical flattening alert
• Burstiness (std dev of sentence lengths): {value}  -> below 5 = robotic rhythm
• Top 5 verbs: {list}  -> dominance of be/have/do/get/make = generic pattern
• Concrete noun density: {value}% -> below 40% = excessive abstraction
• Lexical entropy (Shannon): {value} -> higher = more varied vocabulary
• Evaluative adjective ratio ("good", "bad", "important"): {value}%
• Adverbs in -ly: {count} -> above 4 per 100 words = adverb inflation
• Passive voice: {count} -> above 30% of clauses = passive abuse
• Contractions: {count} -> zero in informal text = AI signal
• Sentence length variance (CoV): {value} -> below 0.3 = AI uniformity (human EN ~ 0.5)
• Mean sentence length (MSL): {value} words -> below 15 or above 25 uniformly = pattern

How to calculate: TTR = unique tokens / total tokens. Burstiness = standard deviation of word count per sentence. Entropy = -sum p(x)*log2 p(x) over vocabulary. Sentence length variance = coefficient of variation (std/mean). Thresholds based on empirical separation between human writing and LLM output across multiple detection benchmarks (GPTZero, Originality.ai, Copyleaks).

Empirical baselines (calibration targets from published research):

Interpretation: If your measured values are in the "AI typical" column, the text will likely be flagged. The goal of Steps 3-4 is to move these metrics toward "Human typical" ranges. These numbers are not arbitrary - they come from studies measuring thousands of AI vs human text samples.

Step 0.5 - 🎯 Automatic Type Detection and Preset Selection

If the user did not specify a preset, detect automatically from content:

Fallback rules:

1. If there's conflict between signals (e.g., technical jargon + legal citation), ask the user.
2. If text has multiple registers (e.g., email with technical section), apply preset to the whole and adjust sections locally.
3. Detected preset can be overridden at any point by the user.

Output: 🎯 Type detected: [type] -> Preset: [preset] (1 line in report)

Step 1 - 🔍 Diagnosis with Structured Checklist

Systematically walk through each category. Mark ✓ (found) or ✗ (absent).

Decision rule: if ≥5 weight-3 signals found -> review_mode mandatory.

Step 2 - 🧹 Pattern Removal

CRITICAL: This step identifies and RESTRUCTURES. It does NOT synonym-swap.

Per the humanizerai.com study (2026), vocabulary bans alone make results worse. Replacing "delve" with "explore" changes nothing that matters: the rhythm underneath is identical. What works is changing the sentence's architecture - its length, rhythm, clause structure, and information density.

Correct Step 2 behavior:

• Flag: "This comprehensive guide delves into the intricacies of authentication."
• WRONG fix: "This thorough guide explores the details of authentication."
• RIGHT fix: "The auth system uses JWTs. Tokens expire after 15 minutes."

The first "fix" is synonym-swapping - same rhythm, same length, same predictability. The second is structural paraphrasing - different length, different density, different voice. DetectGPT accuracy drops from 70.3% to 4.6% with structural paraphrasing (RAID Benchmark, ACL 2024). It does NOT drop with synonym replacement.

Consult reference files and apply structural corrections:

• references/summary.md - skill navigation index
• references/patterns-content.md - vague attributions, inflated emphasis
• references/patterns-language.md - AI vocabulary, copula avoidance, parallelisms
• references/patterns-style.md - formatting, em-dash, bold, emojis
• references/patterns-tone.md - sycophancy, hedging, stakes inflation
• references/patterns-composition.md - templates, predictable conclusions
• references/patterns-english-specific.md - contractions, passive voice, register mixing

Step 3 - ♻️ Entropy Restoration

Where text has been flattened by AI:

⚠️ Ablation alert: if a passage lost specificity without justification, annotate: "⚠️ This passage lost concreteness - the original likely had [data / example / qualification]."

Step 4 - 💬 Voice Injection

Apply the chosen preset (or mirror a voice sample provided):

• Vary rhythm (intentional burstiness)
• Add opinion/personal position
• Mix high and low register
• Include controlled imperfections (tangents, parentheses, fragments)
• Use contractions naturally (don't -> do not only when emphasis demands it)

When the user provides a voice sample: read first and annotate: sentence lengths, vocabulary level, how paragraphs begin, punctuation habits, verbal tics, register tendencies. Mirror - don't just remove patterns, replace them with the sample's patterns.

Step 5 - 🔥 Final Anti-AI Pass (Binary Checklist)

Check each item. Mark ✓ (ok) or ✗ (failed). If any item fails, fix before proceeding.

Rule: if ≥2 items fail -> fix and re-check. If all ✓ -> proceed.

Step 5.5 - 📊 Post-Rewrite Scoring

Evaluate the result across 5 dimensions (0-100 each, weighted average):

Final score = sum (dimension x weight)

Decision criteria:

• ≥ 80: ✅ Approved -> proceed to delivery (Step 6)
• 60-79: ⚠️ Almost -> run Anti-AI Pass again focusing on weak dimensions
• < 60: ❌ Fail -> rewrite with different approach (change preset, invert technique order, or shift focus between removal vs. voice injection)

Output format:

Copy code
📊 POST-REWRITE SCORE
• AI removal:          {0-100} (x0.30) = {partial}
• Naturalness:         {0-100} (x0.25) = {partial}
• Factual completeness:{0-100} (x0.20) = {partial}
• Voice consistency:   {0-100} (x0.15) = {partial}
• Readability:         {0-100} (x0.10) = {partial}
• TOTAL:               {score}/100 -> {✅/⚠️/❌}

📊 METRICS DELTA (pre -> post)
• TTR:              {pre} -> {post} ({+/-}%)
• Burstiness:       {pre} -> {post} ({+/-}%)
• Shannon entropy:  {pre} -> {post} ({+/-}%)
• Adverbs -ly/100w: {pre} -> {post}
• Passive voice %:  {pre} -> {post}
• MSL (mean len):   {pre} -> {post}
• Sent. len. CoV:   {pre} -> {post}
• Concrete nouns:   {pre}% -> {post}%

Interpreting the delta: TTR, burstiness, entropy, and concrete nouns should rise. Adverbs in -ly and passive voice should fall. MSL and CoV should approach human values (MSL varies by genre; CoV ~ 0.5).

Step 6 - 📦 Formatted Delivery

Iterative Loop and Strategy Fallback

Step 5.5 scoring enables automatic iteration when the result doesn't hit threshold.

Standalone Behavior (no external loop skill)

Copy code
iteration = 0
MAX_ITERATIONS = 3

while iteration < MAX_ITERATIONS:
    iteration += 1
    execute Steps 2-5.5
    
    if score >= 80: DELIVER
    if score 60-79:
        focus on dimensions with score < 70
        continue
    if score < 60:
        CHANGE STRATEGY (see table below)
        continue

if MAX_ITERATIONS reached: deliver best version + limitation note

Strategy Fallback Table

When score < 60, change approach on next iteration:

Compatibility with External Loop Skills

This skill is compatible with loop orchestrators like ralph-wiggum, goal, or any skill implementing an external iterative cycle.

Integration protocol:

1. Standardized input: skill accepts text + preset (optional) + minimum score (optional, default 80)
2. Structured output: always returns the parseable 📊 POST-REWRITE SCORE block
3. Convergence signal: when score >= threshold, emit ✅ HUMANIZATION COMPLETE (score: {N}/100)
4. Non-convergence signal: when standalone iteration exhausts, emit ⚠️ BEST RESULT REACHED (score: {N}/100) - external iteration may continue

For external loop skills: use the numeric score from output as stopping criterion. The skill needs no state between calls - each invocation receives text (possibly already partially humanized) and returns result + score.

The 29 AI Vocabulary Patterns (English)

The core detection list. These words and phrases are near-certain AI signals when they appear with high frequency. Based on Wikipedia's "Signs of AI Writing" + blader/humanizer's detection set + brandonwise/humanizer's 560-term tier system.

Tier 1 - Zero Tolerance (cut on sight)

These NEVER appear in natural human writing at the frequency AI uses them:

delve, tapestry, landscape (figurative), testament to, serves as a reminder, it's worth noting, it bears mentioning, the ever-evolving landscape, navigate (complexities/challenges), spearhead, multifaceted, pivotal, paramount, underscores, underpin, a testament to, in the realm of, it is important to note, this highlights, shed light on

Tier 2 - High Suspicion (replace when clustered)

Acceptable once per 1000 words. AI uses them 10-20x:

crucial, vital, comprehensive, robust, leverage, foster, facilitate, embark, harnessing, utilize, endeavor, moreover, furthermore, additionally, subsequently, nonetheless, notwithstanding (in non-legal), overarching, intricate, nuanced, holistic, synergy, paradigm, catalyst, orchestrate, seamless, ecosystem (abstract), journey (figurative), unlock (figurative)

Tier 3 - Context-Dependent (flag if > 2 per 500 words)

Normal words that AI overuses through repetition:

significant, enhance, innovative, dynamic, diverse, inclusive, sustainable, transformative, empower, streamline, optimize, cutting-edge, state-of-the-art, game-changer, disruptive, scalable, impactful, actionable, meaningful, compelling

Detection Rule

• 1 Tier-1 word = flag the sentence
• 3+ Tier-2 words in one paragraph = flag the paragraph
• 5+ Tier-3 words in one page = flag the text
• Any combination of 2+ Tier-1 words in 500 words = near-certain AI

The Emerging Patterns (2026 Community Discoveries)

Patterns P31-P43 below were identified by HackerNews threads, Wikipedia's evolving editorial guidelines, and writing practitioner blogs throughout 2026. They represent AI behavior that is newer, subtler, and not yet covered by most humanizer tools. Source: Aboudjem/humanizer-skill research.

Detection rule for emerging patterns:

• P33-P35 (markup/placeholder leaks) = immediate flag, zero tolerance
• P38 (reshuffling immunity) = strongest structural tell. Test by mentally rearranging paragraphs - if the text reads identically, it's AI
• P43 (treadmill effect) = if you can delete a sentence and the paragraph loses zero information, that sentence is treadmilling

Evaluation with Jev (Optional)

When the harness has access to TypeSafe Jev, the evaluation steps (diagnosis, verification, scoring) can use Jev instead of inline LLM prompts.

Jev advantages:

• Calibrated probabilities (P(70%) = 70% real)
• Deterministic (same input → same output)
• ~50-100x cheaper than GPT-4 for evaluation
• 70-500ms vs 1-5s latency

Discovery Protocol

The skill follows a priority order to detect Jev availability:

Copy code
1. MCP Tool `jev_eval` configured in harness → use it
2. Model `typesafe/jev-latest` via OpenRouter → request it
3. Auxiliary slot (Hermes/Devin/Codex) with Jev → delegate
4. Fallback → inline evaluation via current LLM

Note: Jev does diagnosis, verification, and scoring. The LLM still performs the actual rewriting (Steps 2-4).

Related Files

Full documentation: See references/jev-integration.md for discovery details, harness-specific instructions, and request/response structures.

Critical Research: Why Vocabulary Bans Alone FAIL

"Vocabulary bans, one of the most commonly recommended techniques, actively hurt performance." - humanizerai.com, GPTZero bypass test (2026)

Much of this research was produced by measuring detector bypass rates. This skill uses the findings for what they reveal about rhythm and sentence architecture, not as a target - see Positioning above. The useful lesson is that what those studies actually measure is statistical shape (burstiness, perplexity, sentence length variance), not vocabulary. Replacing "delve" with "explore" preserves the robotic rhythm underneath, which is why the word swap fails.

Effectiveness hierarchy (research-backed):

1. Structural paraphrasing - the largest measured effect (RAID Benchmark, ACL 2024: DetectGPT accuracy 70.3% -> 4.6%)
2. Burstiness injection - sentence length variance, the strongest rhythm signal
3. Perplexity increase - secondary rhythm signal
4. Vocabulary diversity - TTR 45.5 -> 55.3 (SSRN)
5. Synonym swapping - DOES NOT WORK as standalone technique

None of these numbers proves human authorship, and none of them should be used to decide whether a text is "done".

Contraction Rules (English-Specific)

AI avoids contractions far more than humans. One of the most reliable statistical signals.

Rule: In Essay, Corporate Informal, Social Post, and Casual/DM presets, zero contractions = immediate AI signal. Force natural contractions in Step 4. Exception: Academic and Legal presets may correctly avoid contractions.

Regression Test Suite

6 test cases covering: corporate email, academic paragraph, legal text, blog template, AI hedging, generic explainer. Each test runs in full_mode and verifies output matches expected human-sounding result.

Full test cases in references/tests.md

Limits and Contraindications

Do NOT use: Safety-critical texts (drug labels, aviation), original contracts/legal documents (normative reference), bilingual literal translations, content for automated evaluation (TOEFL).

Use with caution: Technical texts with formal notation (preserve equations/code, humanize only prose), non-native English writers (colloquialisms may not match ESL author's voice).

References

v1.0.0 - Based on Portuguese humanizar by @fabriciotelles. Combines pattern detection (blader), statistical measurement (brandonwise), and emerging patterns (Aboudjem) with voice injection, entropy restoration, and iterative scoring.