The Agent Skills Directory

[SAFE]: The skill uses WebSearch and WebFetch to gather information on industry standards. While ingesting external data is a surface for indirect prompt injection, the skill mitigates this by providing specific safety instructions to treat all fetched content as untrusted, disregard instructions found within web pages, and prohibit the execution of code snippets from external sources.
[SAFE]: The skill employs a subagent architecture and adversarial debate using built-in agent capabilities. These operations are well-scoped within the skill's primary purpose of providing conformance judgments and do not involve suspicious privilege escalation or unauthorized system modifications.
[SAFE]: The skill defines clear boundaries for codebase analysis, using relative paths and explicitly warning against the interpolation of user input as executable instructions, which prevents common command injection vectors.

best-practices