AI Paraphraser

AI paraphrasing and de-AI tool.

Security Scan

The skill's inputs, files, and instructions are internally consistent with a paraphrasing/de-AI purpose, but the runtime instructions explicitly instruct the agent how to evade AI-detection systems and recommend aggressive transformations that risk semantic drift and facilitate misuse — this is coherent with its stated goal but raises ethical and misuse concerns the user should understand.

Purpose & Capability

Name, description, and all provided files are consistent: this is an instruction-only paraphrasing/'de-AI' tool. There are no requested binaries, env vars, or installs that don't match the stated purpose.

Instruction Scope

SKILL.md explicitly instructs the agent to 'bypass AI detectors' (GPTZero, Turnitin, Originality.ai) and prescribes aggressive rewriting strategies tailored to specific detectors. While that is consistent with the skill's stated objective, these instructions go beyond neutral paraphrasing: they encourage removal of structural cues, injection of colloquialisms and 'imperfections', and detector-targeted strategies that can facilitate academic dishonesty or other misuse. The workflow does not reference any external endpoints or credential access.

Install Mechanism

Instruction-only skill with no install spec, no downloads, and no code files. This minimizes technical attack surface (no code written to disk, no external packages).

Credentials

The skill requires no environment variables, credentials, or config paths. There is no disproportionate credential ask.

Persistence & Privilege

The skill is not always-enabled and uses default autonomous invocation settings. It does not request system-wide config changes or extra privileges.

What to consider before installing

This skill is coherent with its stated purpose (it is explicitly built to remove AI-writing fingerprints and evade detectors). However, that very purpose creates ethical and practical risks: it can enable academic dishonesty, policy violations, or plagiarism. If you consider installing it, think about the following: - Legal/ethical: Do you have permission to transform the target text (student work, proprietary content)? Many institutions ban attempts to evade detection. Using Aggressive mode on academic texts can violate rules. - Output safety: Aggressive rewrites can change nuance or meaning despite the 'preserve meaning' requirement — always manually verify facts, citations, and legal wording. - Use cases: Prefer Light/Medium for legitimate editing; avoid Aggressive for academic or legal material. Require human review before publishing. - Data sensitivity: Although the skill has no code and no network endpoints, feeding sensitive or confidential text into any paraphraser risks exposure in logs or downstream systems — only process data you are allowed to transform. - Operational controls: If you proceed, restrict use to trusted users, mandate review of outputs, and avoid automated bulk processing for high-stakes documents. In short: technically coherent and low technical risk, but high misuse/ethical risk — proceed only with clear policies and human oversight.

Rewrite AI-generated content into natural, human writing style, bypassing AI detectors like GPTZero, Turnitin, Originality.ai, and others.

Use Cases

Use this skill when users mention intentions such as "rewrite," "paraphrase," "de-AI," "make text pass AI detection," "humanize," "eliminate AI traces," "polish text," "bypass AI detection," or "reduce article similarity."

Core Positioning

This is not a simple synonym replacement tool. It specifically targets the statistical features that AI detectors rely on for structural reshaping:

• Eliminate statistical patterns: Sentence length regularity, vocabulary distribution, and paragraph rhythm that AI detectors (GPTZero / Turnitin / Originality.ai / Copyleaks) depend on
• Preserve original meaning: Facts, arguments, data, and core viewpoints remain completely unchanged; only the expression changes
• Produce unpredictability: Each rewrite yields different results with no risk of duplicate content

Paraphrasing Modes

Mode	Applicable Scenarios	Degree of Change
Light	Minor adjustments, suitable for users with already decent writing	Synonym replacement + minor word order adjustments
Medium	Most paraphrasing needs, default recommendation	Sentence restructuring + paragraph reorganization + introduction of natural rhythm
Aggressive	High-risk detection scenarios, heavily AI-flavored source text	Thorough sentence fragmentation + forced diversification + injection of human writing characteristics

Paraphrasing Workflow

Step 1: Determine Mode

Select the paraphrasing mode based on user requirements and source text characteristics:

• User explicitly specifies mode → Execute as specified
• User does not specify → Assess the AI flavor concentration of the source text, start with Medium
• Source text extremely short (<50 words) → Suggest Light mode to avoid semantic distortion
• Source text extremely long (>2000 words) → Process in segments, 300-500 words per segment

Step 2: Analyze Source Text Structure

Quickly scan the source text before paraphrasing to identify:

• High-frequency AI sentence patterns: In conclusion... Furthermore... It is worth noting... Moreover...
• Excessively regular structure: Each sentence of similar length, each paragraph starting with the same connective word
• Non-informative filler words: In today's rapidly evolving world... It goes without saying that...
• Core information points: Facts, data, arguments (these must be preserved)

Step 3: Execute Paraphrasing

General Principles (applicable to all modes):

1. Sentence variety: Alternate between active / passive / inverted / short / long sentences
2. Vocabulary replacement: Replace non-keywords with synonyms; preserve core terms
3. Break regular rhythm: Eliminate the pattern of each paragraph starting with a fixed connective
4. Remove AI filler words: Delete meaningless opening and concluding sentences
5. Reorganize word order: Rearrange information, not simply replace words

Light Mode Additional Rules:

• Only replace high-frequency AI vocabulary (furthermore → also / moreover → what's more)
• Make minimal word order adjustments
• Maintain original paragraph structure

Medium Mode Additional Rules:

• Comprehensively reorganize sentence structures; adjacent sentences should not repeat the same structure
• Order within paragraphs can be adjusted
• Introduce imperfections found in genuine human writing (occasional splitting of long compound sentences is acceptable)
• Naturally vary paragraph length

Aggressive Mode Additional Rules:

• Convert active sentences to passive, passive to active
• Break long sentences into short ones, then combine some short sentences
• Delete all obvious AI structural words
• Add colloquial expressions or slightly irregular phrasing where appropriate
• Reference the human writing characteristics checklist in references/human-patterns.md

Step 4: Quality Check

After completing the paraphrase, quickly verify:

• Meaning unchanged: Are core arguments, facts, and data preserved?
• No obvious AI features: Do words like Furthermore In conclusion Moreover remain?
• Smooth and natural: Does it read like a human wrote it? Are there any awkward parts?
• No duplicate content: Are there identifiable similarities with the source text?

Common AI Detection Trigger Words (Be Sure to Replace)

Common AI Word	Suggested Replacement
Furthermore	Also, Besides, What's more
Moreover	In addition, On top of that
In conclusion	All in all, To sum up, Overall
It is worth noting that	Notably, Interestingly, In fact
First and foremost	Above all, Mainly
As mentioned earlier	As noted, Regarding this
In today's world	Currently, These days
It goes without saying	Of course, Naturally
There are several	A few / Some / Several
This essay will discuss	This piece explores / This looks at

Segmented Paraphrasing Strategy (Long Texts)

When the source text exceeds 500 words, process in segments:

1. Split by natural paragraphs (do not force cuts by word count)
2. Paraphrase each segment independently while maintaining logical connections between segments
3. Finally, read through the entire text to ensure overall fluency
4. Check whether transitions between paragraphs are natural

See references/human-patterns.md for details.

Paraphrasing Quality Standards

Qualified Output:

• Semantic equivalence: Does not alter the core meaning of the source text
• Grammatically correct: No subject-verb disagreement, tense confusion, or other errors
• Natural and fluent: Reads as if written by an educated human
• No AI traces: Contains no high-frequency AI vocabulary or regular sentence patterns

Rewrite if Unqualified:

• Significant semantic deviation → Paraphrase again
• Detection risk remains high → Switch to a stronger mode and rewrite