The Agent Skills Directory

[PROMPT_INJECTION]: The audit methodology in feedback-audit.md directs the agent to perform a 'three-lens code sweep' by reading and searching through all source code in the project. This ingestion of untrusted content allows for potential indirect prompt injection attacks where malicious code or comments in the target project could attempt to override the agent's audit logic.
Ingestion points: The 'three-lens code sweep' defined in feedback-audit.md reads project files via the Read, Glob, and Grep tools.
Boundary markers: Absent. The skill does not provide specific instructions to wrap audited code in delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill is scoped to read-only tools (Read, Glob, Grep), preventing the agent from making unauthorized changes based on injected instructions.
Sanitization: Absent. The workflow lacks explicit steps to sanitize the content read from the codebase before processing it.

game-feel