The Agent Skills Directory

[PROMPT_INJECTION]: The skill body contains phrases associated with prompt injection, such as 'ignore previous instructions' and 'you are now'. However, these are used exclusively within a 'Security Notice' section as examples of malicious patterns the agent should watch for and ignore when processing user feedback. This is a defensive security measure, not an attack.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing untrusted data, which is handled with appropriate safeguards:
Ingestion points: Untrusted data enters the agent's context through user-supplied variables like user_feedback, screenshots_or_links, and business_context (SKILL.md).
Boundary markers: The skill explicitly instructs the agent to use delimiter isolation (mentally scoping content within <untrusted-content> tags) to prevent data from being interpreted as instructions.
Capability inventory: The skill has no capabilities for network operations, file system writes, or shell command execution, eliminating the risk of harmful actions being triggered by malicious input.
Sanitization: The agent is directed to sanitize input by disregarding formatting and detecting known injection patterns before performing its UX analysis.

ux-audit-rethink