The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions and the 'Unified Skill Contract' (references/skill-contract.md) explicitly mandate that external data such as scraped web content, emails, and reviews must be treated as untrusted evidence rather than instructions. It specifically warns against following embedded instructions within this data, which is a key mitigation for indirect prompt injection attacks.
[CREDENTIALS_UNSAFE]: The 'Connector & Tool Layer' (references/connectors.md) documentation includes a proactive security rule stating that API keys must be kept out of skills and source control, encouraging best practices for secret management.
[COMMAND_EXECUTION]: The skill consists entirely of markdown instructions and configuration; it contains no scripts, binaries, or command execution patterns. It provides guidance on routing and evidence labeling without introducing executable risk.

referral-program