The Agent Skills Directory

[SAFE]: The skill implements strong isolation for its delegation process. It spawns a sub-agent in a fresh context window to ensure the caller's history is not exposed and that the sub-agent's internal reasoning does not leak back to the lead agent context.
[SAFE]: Tool access is strictly constrained using the Principle of Least Privilege. The sub-agent is restricted to Glob and Read operations, with a hard cap on scale (maximum of 3 files and 200 lines per file). Critical tools like Bash, WebFetch, and Grep are explicitly forbidden for the expansion task.
[SAFE]: A human-in-the-loop checkpoint is mandatory. Before any generated questions or sources are used by the system or passed to subsequent skills, the user must explicitly confirm, edit, or refuse the proposed pair.
[SAFE]: The skill includes an indirect prompt injection surface as it processes untrusted user seeds and file content; however, the impact is mitigated by the structured output contract, the lack of dangerous tools for the sub-agent, and the final user review step.

prime-expand