B — Some risk signals Graded October 16, 2026Minor concerns worth a quick check.Based on public adoption and transparency signals as of October 16, 2026. Our opinion, not a code audit or security guarantee.
AI agent skill · General
im-not-ai
AI가 쓴 한글을 사람 글처럼 윤문하는 Claude 스킬 — Korean AI-text humanizer: detects and rewrites translationese, mechanical parallelism, and 71 other AI tells
Bottom line: Signal-based grade (4/7 points): very widely starred on GitHub, public source code. This reflects adoption and transparency, not a code audit.
- Popularity score
- 43/100
- Rank
- #254 of 500
- GitHub stars
- 6k
What it does
Detects and rewrites Korean text generated by AI to sound more natural and human-written, addressing 71+ AI writing patterns including translationese, mechanical parallelism, and repetitive structures. Integrates with Claude as a skill to polish AI-generated Korean content into idiomatic, fluent prose.
Who it’s for
Korean-language content teams, multilingual AI applications, and organizations publishing Claude-generated Korean text that needs to pass as naturally written. Particularly useful for marketing, editorial, and customer-facing Korean communication.
Copy-paste setup
Check these steps against the project’s own instructions before running anything.
Verify against the repo README at https://github.com/epoko77-ai/im-not-ai (opens external site) for Claude integration steps. Likely implementation via Claude custom instructions or prompt injection, though specific setup method is unclear from available signals. Check repository for Claude Desktop mcpServers configuration if available.
Risks
- Prompt injection via humanization: Rewriting rules may inadvertently execute or amplify hidden instructions embedded in AI-generated text.
- Data processing: Korean text content is sent through the skill's processing pipeline—verify privacy policy for retention or logging practices.
- False confidence: Successfully humanized text may mask factual errors or hallucinations, increasing risk of undetected misinformation in published content.
- Supply-chain reliance: Dependency on a single-maintainer GitHub project with no package distribution signal (0 npm downloads) creates maintenance continuity risk.
- Language-specific brittleness: Specialized Korean NLP may fail gracefully or produce unpredictable results on mixed-language or edge-case text.
How to stay secure
Least-privilege permissions
- Limit user access to only those who genuinely need the humanization skill; don't enable it organization-wide by default
- Restrict input scope to already-vetted AI-generated content (like internal drafts), not external or untrusted text sources
- Disable auto-apply features if available—require manual review before the skill rewrites any text, so humans stay in control
- Separate credentials if the skill connects to other systems; never reuse passwords across your main accounts and this tool
Network isolation
- Route through a sandbox or isolated testing environment before applying the skill to production content; don't let it process sensitive customer data directly
- Monitor outbound traffic to confirm Korean text is only sent to the skill's service and nowhere else unexpectedly
- Verify HTTPS encryption is enforced; check that connections use secure protocols, not unencrypted data transmission
- Log all API calls to the skill with timestamps and input/output samples so you can audit what was processed
Monitoring
- Track humanization changes by comparing original text to rewritten output regularly; flag any suspicious new phrases or injected instructions that shouldn't be there
- Set alerts if the skill suddenly fails, crashes, or produces garbled output (sign of a breach or abandonment)
- Review published content before distribution—spot-check humanized text for factual accuracy and tone, don't assume the rewrite is error-free
- Check the GitHub project monthly for security updates or signs the maintainer has gone inactive (no commits in 6+ months = high risk)
Red flags to watch for
- Unexpected rewording that changes meaning, adds marketing language, or inserts subtle calls-to-action that weren't in the original
- Longer processing times or intermittent timeouts, suggesting the service is unstable or compromised
- Humanized text that sounds too polished or contains industry jargon you didn't input—possible sign of prompt injection working backwards
- No update or deprecation notice from the maintainer after 12+ months; tool may be unsupported and vulnerable to new attacks
Alternatives
- Grammarly or language-specific editors – Native Korean text polish without AI-detection framing; broader tool maturity.
- OpenAI's text-davinci or GPT-4 fine-tuning – General-purpose humanization via prompt engineering without specialized Korean detection.
- Korpora or KoBART-based local models – Open Korean NLP models run locally to avoid external data transfer.
- Manual editorial review + style guides – Human copyediting remains the gold standard for high-stakes Korean content.