The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill takes arbitrary user-provided topics and processes them through a multi-agent workflow. While the prompts use clear delimiters like 【Topic】, there is a risk that malicious instructions within a topic could influence the sub-agents.
Ingestion points: The topic variable is used to initialize the discussion, define experts, and frame each round's debate in SKILL.md.
Boundary markers: Prompt templates for experts and moderators use 【 and 】 as delimiters for context sections.
Capability inventory: The orchestrator utilizes Bash (for directory creation and logging), Write and Read (for state persistence), and the Teammate and Task APIs (for spawning and communicating with sub-agents).
Sanitization: The provided logic mentions escapeForShell for logging progress, but the primary protection against injection is the structural framing of the prompts.
[COMMAND_EXECUTION]: The skill performs shell operations to manage its internal directory structure (~/.claude/discussions/) and to maintain a live progress log. While these are intended behaviors for session management, they represent a capability that could be targeted via injection in the discussion topic.

swarm-discussion