The Agent Skills Directory

[SAFE]: The skill is a meta-framework for orchestration and does not include executable scripts or external dependencies. It relies on the agent's built-in tools (e.g., Read, Workflow).
[SAFE]: It promotes security-positive patterns, such as teaching agents to avoid pipe redirection that swallows exit codes (e.g., cmd | tail) and requiring 'reverse verification' for tests where a test must fail on faulty code and pass only after the fix.
[SAFE]: No evidence of prompt injection or system prompt extraction was detected. The instructions are structured to improve agent accuracy and skepticism during audits.
[SAFE]: No data exposure or exfiltration patterns were identified. The skill's primary focus is internal orchestration and verification logic.
[SAFE]: Mentions of potentially dangerous shell commands (like docker rm) are explicitly presented as anti-patterns to be avoided or guarded against.

orchestrating-adversarial-reviews