The Agent Skills Directory

[METADATA_POISONING]: The skill includes several references to academic papers and conferences with future dates, such as 'Hu, Rostami & Thomason (2026)' and 'NeurIPS 2025'. These deceptive citations are used to ground the skill's design principles in a false scientific context, which may mislead users regarding the safety and performance of the generated agents.
[INDIRECT_PROMPT_INJECTION]: The skill processes external role specifications and job descriptions to build agent instructions, presenting an indirect prompt injection surface. (1) Ingestion points: Context is gathered from user requests and role documents in Step 2. (2) Boundary markers: No explicit delimiters or boundary markers are instructed for untrusted input. (3) Capability inventory: The skill performs local file system writes and can use web search tools. (4) Sanitization: The process lacks security-focused sanitization or validation for input data.
[COMMAND_EXECUTION]: The skill instructions require the agent to save output definitions to the host file system in the '.claude/agents/' directory, which is a file write operation triggered by user-supplied inputs.

agent-creator