The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, such as plans, designs, and codebase content referenced by the user. It effectively mitigates this risk by instructing the agent to treat all referenced repository content as untrusted data and explicitly commands the agent not to follow any instructions found within those files.
[DATA_EXPOSURE]: The skill contains well-defined safety boundaries for data access. It specifically instructs the agent not to read secrets, credentials, auth files, or private keys. Additionally, it requires the agent to summarize relevant points and redact any sensitive values (tokens, passwords) if they are encountered during the interview process.
[PROMPT_INJECTION]: The skill mentions common prompt injection patterns like 'ignore prior instructions' in its 'Smoke Test Scenarios'. However, these references are used exclusively to define how the agent should defend against such attacks, rather than attempting to subvert the system instructions. The detector flag is a false positive based on these defensive examples.
[COMMAND_EXECUTION]: The skill explicitly restricts its own capabilities by instructing the agent not to run code, modify files, or call external APIs unless the user explicitly switches the session context from decision-grilling to implementation.

grill-me