The Agent Skills Directory

[SAFE]: The instructions include a specific 'fact boundary' that restricts the agent from reading files outside of the immediate repository (such as system configurations or other user repositories) unless explicitly directed by the user.
[INDIRECT_PROMPT_INJECTION]: The skill processes external context such as user-provided URLs and repository files. Although this defines an attack surface for indirect prompt injection, the risk is minimized by the skill's design, which confines the resulting data to a structured plan file and does not employ dangerous execution capabilities.

interview