The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to detect (command -v) and execute local AI CLI agents such as codex, claude, gemini, and cursor-agent. The skill explicitly warns about the risks of agents that lack native read-only sandboxing (specifically cursor-agent) and recommends user consent before use.
[EXTERNAL_DOWNLOADS]: The skill's primary function involves sending code snippets and plans to external AI providers for review. It includes explicit 'Guardrails' and 'Anti-patterns' sections instructing the user and the agent to redact credentials, secret-bearing paths, and sensitive environmental data before transmission.
[PROMPT_INJECTION]: Static analysis flagged an instruction-override pattern in resources/brief-template.md. Analysis confirms this is a defensive security measure: the skill instructs the verifier agent to explicitly look for and report hidden instructions in the code it reviews (indirect prompt injection defense), rather than a malicious attempt to override the host's behavior.
[SAFE]: The skill implements strong boundary markers by feeding review briefs via stdin rather than shell arguments, preventing brief content from leaking into process listings or shell history.

double-check