The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data from website interactions, creating a surface for indirect prompt injection.
Ingestion points: branchly_read_sessions, branchly_read_session_detail, and branchly_get_top_searches (referenced in SKILL.md) retrieve raw user queries and conversation transcripts from the platform.
Boundary markers: The skill includes a "Pitfalls to Avoid" section that explicitly directs the agent to filter out and ignore prompt injections or off-topic noise found within the processed data.
Capability inventory: The skill is scoped as read-only and analytical; it does not contain tools for command execution, file writing, or external network operations beyond the provided MCP server tools.
Sanitization: The skill relies on model-level filtering instructions to identify and discard malicious or irrelevant content during the recommendation workflow.

content-ideas