The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted agent definition files (e.g., agents/<name>.md) which could contain malicious instructions designed to influence the agent's behavior during analysis.
Ingestion points: The skill ingests text from contributor-authored files via Read, Grep, and Glob tools.
Capability inventory: The skill is restricted to read-only analysis tools, preventing it from executing commands, writing files, or performing network operations that a successful injection might attempt to trigger.
Boundary markers: The skill includes specific instructions to treat the material under review strictly as data and to ignore any instructions found within it, explicitly flagging attempts to direct the review as a security violation.
[NO_CODE]: The skill consists exclusively of markdown instructions and YAML frontmatter configuration, with no accompanying scripts or binary executables.

reviewing-agent-definitions