The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input and has access to file manipulation tools, which presents a surface for indirect prompt injection attacks.
Ingestion points: The agent ingests user-supplied text for processing in the SKILL.md instructions.
Boundary markers: There are no specific delimiters (like triple-backticks or unique tags) or instructions to the agent to ignore instructions embedded within the user-provided text.
Capability inventory: The skill is granted Read, Write, Edit, Grep, and Glob tools in the frontmatter of SKILL.md.
Sanitization: The instructions do not include any steps to sanitize or validate input text before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The installation guide references an external code repository.
Evidence: The README.md instructs users to download the skill using git clone https://github.com/blader/humanizer.git. While this is a common practice for installation, it involves downloading content from an external source.

humanizer