The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process user-provided text, which is an untrusted data ingestion surface. This creates a potential vector for indirect prompt injection where a user might embed hidden instructions within the text to be 'humanized'. However, the skill provides specific guardrails and a verification loop to mitigate this risk.
Ingestion points: The skill accepts arbitrary text strings from the user for stylistic editing in SKILL.md.
Boundary markers: There are no explicit delimiters (like XML tags) used for the input text, which increases the risk of the model confusing data with instructions.
Capability inventory: The skill is restricted to text transformation and does not have access to file system writes, network operations, or subprocess execution.
Sanitization: The skill employs a 'Self-Verification Checklist' (자가검증 체크리스트) and an 'Audit' phase. These instructions force the agent to review its own output for fidelity to the original meaning and to ensure it hasn't introduced its own artificial patterns, which acts as a secondary defense against malicious input override.
[EXTERNAL_DOWNLOADS]: The skill mentions an optional external skill korean-character-count for precise character counting. This is a common pattern for modular skill usage within agent platforms and does not constitute a security risk. It also references several legitimate open-source GitHub repositories (epoko77-ai/im-not-ai, blader/humanizer) for academic and technical attribution, which is standard practice for MIT-licensed software.

korean-humanizer