The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references documentation and configuration examples from external GitHub repositories and official Claude documentation. These references are purely informative and do not involve the automated download or execution of untrusted scripts.
[COMMAND_EXECUTION]: While the skill documents that subagents can use tools like Bash and Edit, it emphasizes strict tool configuration and the principle of least privilege. It explicitly warns against 'permission sprawl' and recommends granting only the minimum necessary tools for a subagent's task.
[PROMPT_INJECTION]: The skill includes a dedicated section on prompt injection in its debugging and security references. It provides clear defensive patterns, such as instructing agents to treat user input as data rather than instructions and reinforcing that instructions must only come from the system prompt.
[DATA_EXFILTRATION]: No patterns indicative of sensitive data access or unauthorized exfiltration were found. The skill focuses on project-level orchestration and provides guidance on secure context management.
[REMOTE_CODE_EXECUTION]: There are no instances of remote code execution or fetching executable content from untrusted sources. The execution model described is confined to the Claude Code platform's native subagent primitives.

agent-creator