The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text provided by the user (LinkedIn drafts) to perform humanization and auditing passes. It possesses capabilities to write to the local filesystem (updating references/voice-profile.md) and make network calls to external APIs. This configuration presents a surface for indirect prompt injection, although the specific regex-based scrubbing logic and lack of complex tool chaining mitigate the risk of arbitrary instruction execution.
Ingestion points: User-provided text in SKILL.md (input text for humanization).
Boundary markers: Absent; the skill does not define specific delimiters to wrap untrusted input when processing it.
Capability inventory: File system write access for voice profile persistence (sub-skills/voice-profile.md); network operations to AI detection APIs (scripts/test_detectors.py).
Sanitization: Not explicitly implemented in the prompt instructions.
[COMMAND_EXECUTION]: The skill includes a utility script scripts/test_detectors.py which is designed to be executed by the agent to interface with third-party AI detection APIs. The script processes user input as data for HTTP requests, avoiding unsafe shell execution or interpolation of the untrusted text.
[EXTERNAL_DOWNLOADS]: The skill documentation and scripts reference external AI detection services (including GPTZero, Originality.ai, ZeroGPT, Sapling, and Copyleaks) and image generation tools (Pixfaro). These are categorized as well-known technology services integrated for specific functional purposes and are documented neutrally.

linkedin-humanizer