The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The Experts skill is designed to ingest and assess user-provided content, including code, documents, and problem statements, creating a potential surface for indirect prompt injection where adversarial text could influence the expert panel's reasoning.
Ingestion points: The skill instructions in SKILL.md (Workflow section 1) explicitly guide the agent to extract and restate the 'exact question' and 'relevant code, documents, systems, or artifacts' from the user request as the primary basis for assessment.
Boundary markers: The skill does not define or mandate the use of explicit boundary markers or delimiters (such as XML tags or JSON schemas) to isolate user-provided data from the agent's instructions, nor does it explicitly instruct sub-agents to ignore instructions embedded within those artifacts.
Capability inventory: As no restrictions are defined in the frontmatter's allowed-tools, the skill operates with the agent's full set of default capabilities (including file system access, terminal execution, and network operations), which could be misused if the expert panel process is compromised by malicious input.
Sanitization: The provided instructions do not include mechanisms for sanitizing, validating, or escaping user-supplied context before it is processed by the individual expert personas and synthesized into the final recommendation.

experts