The Agent Skills Directory

[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by instructing the agent to 'Read every candidate end to end' and 'Graft the best ideas from the others into it'. If a subagent processes untrusted external data, it could produce malicious output that influences or redirects the parent agent's behavior during the synthesis phase.\n- Ingestion points: The parent agent reads artifacts and rationales from subagent output paths (e.g., /tmp/arena-<slug>/candidate-<n>/).\n- Boundary markers: The instructions lack explicit requirements for using delimiters or 'ignore instructions' markers when reading and processing subagent outputs.\n- Capability inventory: The skill utilizes subagent orchestration capabilities (run_in_background) and file system access to manage outputs.\n- Sanitization: No sanitization, validation, or escaping of subagent-generated content is mandated before the parent agent incorporates the ideas into the final artifact.

arena