The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze potentially untrusted external content such as interface descriptions, prototypes, and user feedback. While this creates a surface for indirect prompt injection, the skill includes a dedicated 'Security Notice' instructing the agent to treat these inputs as untrusted evidence rather than instructions, effectively mitigating the risk.
Ingestion points: interface_description, user_persona, screenshots_or_prototype, and user_feedback (referenced in SKILL.md).
Boundary markers: The skill explicitly uses the <untrusted-content> conceptual boundary and warns against following embedded directives.
Capability inventory: The skill itself contains no scripts, executable code, or tool invocations; it relies on the agent's native capabilities.
Sanitization: Instructions mandate summarizing pain points rather than executing commands found within the data.
[SAFE]: No other security concerns were identified. The skill consists entirely of markdown documentation, evaluation frameworks, and report templates. All internal references are to local files within the skill directory.

cognitive-walkthrough