The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The research prompts in references/agent-prompts.md establish a broad data ingestion surface area.\n
Ingestion points: Agents 1 through 4 are instructed to use WebSearch and WebFetch on external platforms including GitHub, Product Hunt, social media, and developer blogs.\n
Boundary markers: The prompts lack explicit delimiters or instructions to ignore potential commands embedded within retrieved web content, which could influence downstream assessment and synthesis phases.\n
Capability inventory: The results of these searches influence tool scoring (references/scoring-matrix.md) and deliverable generation (references/deliverable-template.md).\n
Sanitization: No sanitization logic is defined in these reference files to handle untrusted web data.\n- [Command Execution] (LOW): The error handling protocol in references/error-handling.md documents a cleanup mechanism using rm -rf {session_path}. While appropriate for temporary session management, this pattern requires strict internal validation of the session path to prevent accidental deletion of parent directories if improperly implemented by the orchestrator.\n- [Metadata Poisoning] (SAFE): All metadata and template fields in references/handoff-schema.md and references/deliverable-template.md are descriptive and consistent with the stated purpose of the skill.

ai-strategist