The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to override standard AI behavior, such as "Never mention you're an AI," "Avoid disclaimers," and "Skip directly to answering" found in references/prompting-and-workflow.md. While these guidelines are intended to help the agent adopt a "human" persona for writing tasks, they represent a prompt injection vector by instructing the model to hide its nature and bypass safety-related transparency messaging.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data without adequate security boundaries, creating a theoretical surface for injection.
- Ingestion points: User requests are ingested via the $ARGUMENTS variable in SKILL.md.
- Boundary markers: Absent. There are no delimiters or "ignore embedded instructions" warnings to separate the user's prose request from the skill's logic.
- Capability inventory: None. The skill package does not include any executable code or tools that could be leveraged for malicious actions like network exfiltration or file system access.
- Sanitization: Absent. The skill does not implement any validation or filtering for the external content it processes.
[NO_CODE]: No executable scripts, binaries, or platform configurations that grant tool access are present. The skill is entirely composed of documentation and instructional text.

human-writing