The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes open-ended user concepts (Step 1) and interpolates this content into prompts for specialized subagents (Step 2). There are no explicit instructions for the agent to sanitize or use specific boundary markers for this untrusted data before subagent invocation. This creates a surface where instructions embedded in the user's concept could influence subagent behavior.\n- Environment Access for Metadata Resolution: The skill instructions direct the agent to resolve the author's identity by executing git config user.name or checking environment variables like GIT_AUTHOR_NAME. This is used to populate the Author field in the generated concepts_brief.md artifact but involves querying the user's local system configuration.\n- Subagent Coordination: The skill utilizes platform-specific tools (invoke_subagent, Agent, spawn_agent) to launch parallel reviews. While these tools are standard for multi-agent workflows, the skill relies on the underlying platform's ability to maintain context isolation across the six specialized Goldfish lenses.

egm-brainstorm