The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external handoff bundles from user-provided URLs or local files. These bundles contain markdown content and HTML prototypes parsed by sub-agents to generate code. This represents a vulnerability surface where malicious instructions embedded in the design bundle could attempt to influence the agent's behavior.\n
Ingestion points: The bundle_input variable, derived from the positional command argument, in Phase 1.\n
Boundary markers: No specific delimiters or instructions are provided to sub-agents to ignore potential commands within the ingested bundle data.\n
Capability inventory: The skill has access to high-privilege tools including Read, Write, Edit, Bash, WebFetch, and Agent.\n
Sanitization: There is no explicit sanitization or validation of the text content within the fetched bundle before LLM processing.\n- [COMMAND_EXECUTION]: The skill performs shell operations on external data. In Phase 1, it instructs a sub-agent to execute a tar extraction command on a file downloaded from an external URL. While functional, this involves executing high-capability system commands on untrusted external input.

design-import