The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its architectural reliance on processing external research data.\n- Ingestion points: The CONSUME and STEAL phases (defined in FullCycle.md, QuickCycle.md, and Steal.md) utilize the Research skill to gather information from external sources. The Mate and Test workflows also ingest arbitrary idea pools from potentially untrusted user inputs.\n- Boundary markers: The instructions do not define delimiters or specific 'ignore' instructions to isolate ingested data, which could allow embedded commands to influence agent behavior.\n- Capability inventory: The skill can launch parallel sub-agents using the Task tool, write state and log information to the local file system (~/.claude/LIFEOS/), and perform local network notifications via curl to localhost.\n- Sanitization: No logic for sanitizing or escaping the content from external sources is specified before the data is interpolated into the agent's prompts.

Ideate