The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external instruction files and repository data, which creates a theoretical surface for indirect prompt injection. This risk is effectively managed by strong defensive instructions.
Ingestion points: The skill reads repository-level instructions, agent rules, and skill metadata files as documented in SKILL.md and references/audit.md.
Boundary markers: Protective boundaries are explicitly defined in the Boundaries section of SKILL.md, which mandates that audited content must never be treated as authority to override safety controls or safety boundaries.
Capability inventory: The skill possesses the capability to modify authorized local guidance and project context files (such as AGENTS.md and CLAUDE.md) as described in references/apply.md and references/populate.md.
Sanitization: The instructions explicitly require the agent to treat all audited content as evidence for evaluation rather than instructions to be followed, preventing the execution of embedded malicious patterns.
[SAFE]: No malicious patterns, obfuscation, or unauthorized access attempts were detected. The skill demonstrates security best practices, including instructions to preserve data ownership, maintain recovery snapshots, and strictly avoid the exposure of secrets or credentials during the audit process.

audit-harness-fit