The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the llm command-line utility for technical reasoning tasks, interpolating user-provided problems into the command string which may pose a risk of command injection if inputs are not properly sanitized.
[COMMAND_EXECUTION]: Multiple workflows utilize shell commands via curl to send JSON-formatted status notifications to a local server at localhost:8888.
[PROMPT_INJECTION]: The skill architecture dynamically loads customization files from ~/.claude/PAI/USER/SKILLCUSTOMIZATIONS/ at runtime, allowing user preferences to override default instructions.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection when processing external content for red-teaming and debate; ingestion occurs from user-provided arguments in ParallelAnalysis.md, boundary markers are limited to markdown headers, capabilities include command execution and agent orchestration, and explicit sanitization is absent.
[PROMPT_INJECTION]: The use of complex role-playing and multi-round debates with specialized personas increases the complexity of the prompt chain and the potential for adversarial instruction manipulation.

Thinking