Humanization

When to Use

Activate when the user wants to:

• Check whether an essay will be flagged as AI-generated
• Reduce the AI-detection score of a piece of writing
• "Humanize" text for academic or journal submission
• Pass AI detection tools (ZeroGPT, GPTZero, etc.)

Overview

This skill feeds an essay to zerogpt.com, reads the AI percentage score, and iteratively rewrites the essay until the score drops below 70%. All rewrites preserve the original meaning and maintain a professional, academic tone appropriate for a learning journal — no casual language, no first-person storytelling, no informal phrasing.

Steps

Phase 1 — Get the Essay

1. If the user has pasted the essay directly in chat, use that text.
2. If the user references a file, use filesystem_read to load it.
3. Store the essay text as the working draft.

Phase 2 — Check AI Score on ZeroGPT

1. Navigate to https://www.zerogpt.com using navigate_page on the current tab or a background tab if the user is on a page they want to keep.
2. Take a take_snapshot to locate the text input area.
3. Clear any existing content in the textarea, then fill it with the working draft.
4. Click the Detect Text (or equivalent submit) button.
5. Wait for the result to appear — use take_snapshot or get_page_content to read the AI percentage score.
6. Extract the numeric score (e.g. "82% AI-generated" → 82).
7. Report the score to the user: "ZeroGPT score: 82% — above the 70% threshold. Beginning rewrite..."

If the score is already below 70%, skip to Phase 4 (report success).

Phase 3 — Rewrite Loop (score ≥ 70%)

Repeat the following loop up to 5 iterations. Stop as soon as the score drops below 70%.

3a. Rewrite the Essay

Apply the humanization strategies below to produce a revised draft. Do not apply all strategies at once — make targeted, incremental changes each round so the meaning is never distorted.

Humanization Strategies (apply progressively)

Round 1 — Sentence-level variation

• Break up long, uniform sentences into shorter ones, or combine short choppy ones into longer compound sentences.
• Vary sentence openings — avoid starting multiple sentences with the same word or structure.
• Replace overly smooth transitional phrases ("Furthermore,", "Additionally,", "In conclusion,") with more varied connectives or restructured paragraphs.

Round 2 — Vocabulary naturalisation

• Replace high-frequency AI vocabulary: "delve", "crucial", "leverage", "utilize", "it is important to note", "in the realm of", "showcasing", "underscoring", "shed light on", "tapestry", "landscape" → use plainer or more specific alternatives.
• Introduce minor lexical variety: synonyms, discipline-specific terminology, or concrete examples where the original was abstract.

Round 3 — Structural reordering

• Reorder supporting points within paragraphs (flip the order of two sub-arguments).
• Split one dense paragraph into two, or merge two thin paragraphs into one.
• Move a concluding sentence to the opening of its paragraph as a topic sentence (or vice versa).

Round 4 — Hedging and epistemic language

• Add appropriate academic hedging where claims are overly definitive: "suggests", "appears to", "may indicate", "tends to", rather than absolute assertions.
• Insert a brief acknowledgement of complexity or nuance where the text is too clean and declarative.

Round 5 — Rhythm and voice refinement

• Introduce one or two concrete, grounded examples or observations (e.g. referencing a specific course concept, a real scenario encountered, or a specific skill practised) to anchor abstract statements.
• Vary paragraph length — ensure not all paragraphs are the same number of sentences.

Tone Guardrails (must follow on every round)

• ❌ Do NOT use casual language ("kinda", "pretty much", "a lot", "stuff")
• ❌ Do NOT add personal anecdotes or first-person emotional statements unless already present in the original
• ❌ Do NOT change technical terms, defined concepts, or proper nouns
• ❌ Do NOT alter the core argument, thesis, or conclusions
• ✅ DO keep an academic, reflective register appropriate for a learning journal
• ✅ DO preserve all citations, references, and section headings if present

3b. Re-check on ZeroGPT

1. Clear the ZeroGPT textarea.
2. fill it with the new draft.
3. Click Detect Text again and read the new score.
4. Report progress: "Round 2 rewrite: ZeroGPT score now 61% ✅ — below threshold."

3c. Decide

• If score \x3C 70% → exit loop, proceed to Phase 4.
• If score ≥ 70% → apply next round of strategies, repeat.
• If 5 rounds are exhausted and score is still ≥ 70% → proceed to Phase 4 with the lowest-scoring draft and report the limitation.

Phase 4 — Deliver Results

1. Present the final rewritten essay in full in the chat (or offer to save it to a file if it is long).
2. Show a summary table:

| Round | AI Score | Status        |
|-------|----------|---------------|
| 0     | 82%      | ❌ Too high   |
| 1     | 74%      | ❌ Too high   |
| 2     | 61%      | ✅ Passed     |

3. If the final score is still ≥ 70% after 5 rounds, be transparent:

   "After 5 rewrite rounds, the lowest score achieved was XX%. The essay has been improved but further manual editing may be needed. The best version is provided below."

4. Optionally offer to save the result: "Would you like me to save the humanized essay to a file?"

Error Handling

• ZeroGPT is slow to load: Wait and retry take_snapshot up to 3 times before reporting a timeout.
• Score not found on page: Use get_page_content to search for percentage text; fall back to evaluate_script to extract it from the DOM.
• ZeroGPT blocks paste/fill: Try evaluate_script to set the textarea value directly, then trigger an input event:

  const ta = document.querySelector('textarea');
  const nativeInputValueSetter = Object.getOwnPropertyDescriptor(window.HTMLTextAreaElement.prototype, 'value').set;
  nativeInputValueSetter.call(ta, `ESSAY_TEXT`);
  ta.dispatchEvent(new Event('input', { bubbles: true }));
  

• Essay is very long (>3000 words): Warn the user that ZeroGPT has input limits. Offer to check the essay in sections and report an average score.

Notes

• This skill works entirely through browser automation — no external APIs required.
• ZeroGPT's scoring can vary slightly between submissions; a score of 68–72% near the threshold may fluctuate. If the score is within 5% of the threshold, flag this to the user.
• The goal is to make the writing sound like a thoughtful human student wrote it — not to strip all structure, but to break the statistical patterns that AI detectors look for.