The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as project source code, documentation, and potentially web research results, to inform its recommendations. This creates a surface for indirect prompt injection where malicious instructions could be embedded in the data being analyzed (e.g., hidden in code comments or markdown files) to bias the brainstorming verdict or influence the agent's behavior.
Ingestion points: SKILL.md (Step 1: "inspect relevant code, docs"), blindspot-pass.md ("research with whatever web tools are reachable").
Boundary markers: The instructions do not explicitly define delimiters or specific "ignore" markers for the external content being processed.
Capability inventory: The skill is designed to read local files, perform web research (where tools are available), and transition to other skills like interview-me and models-consensus. It does not demonstrate file-write or network-exfiltration capabilities in the provided instructions.
Sanitization: No explicit sanitization or filtering of external content is described in the provided logic.

brainstorm