The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because it is designed to ingest, process, and "mentally simulate" instructions from external, untrusted skill files. If an analyzed skill contains malicious prompt injection payloads, the agent performing the analysis could be influenced into unauthorized actions.
Ingestion points: The skill loads SKILL.md and associated files (scripts, examples, references) from a user-specified path provided via $ARGUMENTS in Phase 0.
Boundary markers: There are no explicit instructions or delimiters defined to isolate the content of the analyzed skill from the analyzer's own system prompt or to instruct the agent to ignore embedded instructions during simulation.
Capability inventory: The skill is granted extensive capabilities including Edit, Write, and the Task tool (for invoking other sub-agents), which could be leveraged if the analysis logic is compromised.
Sanitization: No sanitization or validation of the ingested skill content is performed before the "Mental Simulation" (Phase 2a) begins.
[EXTERNAL_DOWNLOADS]: The skill uses WebSearch and WebFetch tools in Phase 2b to verify factual claims against external documentation.
The instructions specify that verification should be performed against official documentation for tools, APIs, and CLI flags.
This behavior is consistent with the skill's primary purpose of ensuring accuracy and is directed toward well-known technical documentation sources.

improve-skill