The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by ingesting data from untrusted external sources like web searches, external documentation via the docs-seeker skill, and visual materials via the ai-multimodal skill. This content is summarized and passed as context to downstream commands without explicit sanitization or boundary markers, which could allow malicious instructions in external content to influence the agent. Findings based on evidence chain: (1) Ingestion points: WebSearch, docs-seeker, and ai-multimodal skills; (2) Boundary markers: Absent; (3) Capability inventory: Database access via psql and command execution via slash commands; (4) Sanitization: None identified.
[COMMAND_EXECUTION]: The skill is configured to execute database queries using the psql command to inspect project structure and data. It also triggers slash commands (/plan:fast, /plan:hard) upon user confirmation, passing the accumulated brainstorm summary as an argument, which creates an automated execution path for potentially influenced data.

brainstorming