The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands to perform operational tasks. It uses curl to send POST requests to a local service (http://localhost:31337/notify) for voice notifications during workflow execution. Additionally, it uses echo to append execution metadata (timestamp, workflow name, input summary, and duration) to a local JSONL file located at ~/.claude/LIFEOS/MEMORY/SKILLS/execution.jsonl.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and decompose user-provided content such as plans, strategies, and arguments. This creates a surface for indirect prompt injection, as malicious instructions could be embedded within the content intended for analysis.
Ingestion points: User input provided for the ParallelAnalysis and AdversarialValidation workflows.
Boundary markers: The skill uses <instructions> tags and structured template blocks to isolate instructions for the various agent personas.
Capability inventory: The agent can execute local shell commands (curl, echo) and interact with other system tools if allowed by the platform configuration.
Sanitization: No specific sanitization or escaping of the user-provided 'argument' is defined before it is interpolated into the agent prompt templates.

RedTeam