The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is explicitly instructed to collect evidence from external, untrusted sources such as creator portfolios, press kits, and museum collections, which may contain malicious instructions designed to influence the agent.
Ingestion points: External URLs and publications specified in the "Collect usable evidence" section of SKILL.md.
Boundary markers: The skill lacks explicit instructions to treat external data as untrusted or to use delimiters to prevent command injection from those sources.
Capability inventory: The agent has the ability to write files to the local file system (within {agentRoot}/research/) and transition to further production rules using the art-direction skill.
Sanitization: No sanitization or validation logic is defined for the content retrieved from external sources.
[DYNAMIC_EXECUTION]: The skill contains instructions that direct the agent to load logic or read its configuration file directly, bypassing platform-provided mechanisms.
Evidence: SKILL.md states: "Load its entrypoint through the harness's skill mechanism, or read SKILL.md directly when no mechanism is available."

visual-research