The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to add a third-party plugin marketplace (duthaho/claudekit-marketplace) and install the claudekit plugin. This behavior downloads code from an unverified GitHub repository.
[REMOTE_CODE_EXECUTION]: Installing a plugin directly into the agent's execution environment allows external code to run with the agent's full permissions, including access to the local filesystem and network tools.
[COMMAND_EXECUTION]: The instructions require running /claudekit:init, which executes opaque logic via the Bash tool to scaffold project files. This command could perform undocumented system modifications or file operations.
[PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection.
Ingestion points: Configuration files generated in .claude/rules/, .claude/modes/, and .claude/hooks/ (SKILL.md).
Boundary markers: Absent. There are no instructions for the agent to validate or treat these generated files as potentially untrusted data.
Capability inventory: The skill requests access to Bash, Read, Write, and WebFetch tools.
Sanitization: Absent. The skill does not implement any content validation for the files it generates or the data it processes during the initialization wizard.

claudekit