The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests untrusted data from local code files and external documentation via WebSearch and WebFetch. However, it includes an explicit boundary stating it is a read-only skill that will not modify code, which significantly limits the impact of potential injections to the explanation output itself.
[COMMAND_EXECUTION]: The skill utilizes tools such as Bash, Grep, and Read to analyze codebases. These are used within the context of a 'read-only explanation skill' to gather source material and trace dependencies, rather than for executing arbitrary or malicious commands.
[DATA_EXPOSURE]: The skill contains instructions for platform compatibility that involve mapping platform-specific directories (e.g., mapping ~/.claude/ to ~/.cursor/skills/). These instructions are limited to configuration path translations for the agent's own metadata and do not target sensitive user credentials, SSH keys, or environment secrets.

explain