The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to read external configuration or context files such as .agents/product-marketing.md, .claude/product-marketing.md, or product-marketing-context.md at the start of a task. This represents an attack surface where instructions embedded in those external files could potentially influence or override the agent's behavior.
Ingestion points: Reads project-level context files including .agents/product-marketing.md, .claude/product-marketing.md, and product-marketing-context.md.
Boundary markers: The instructions do not define clear boundaries (such as XML tags or specific delimiters) or provide "ignore embedded instructions" warnings for the content read from these files.
Capability inventory: The skill itself is instructional and does not define dangerous tools, though it functions within the broader agent environment.
Sanitization: There is no mention of sanitizing, filtering, or validating the content retrieved from these external files before processing.

ab-testing