The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, specifically interview transcripts, recording summaries, and interviewer notes, which creates a potential surface for indirect prompt injection attacks.
Ingestion points: External data such as transcripts or recording summaries provided by the user (described in SKILL.md under "What to ask for").
Boundary markers: The instructions mandate a structured Markdown format and specific handling of quotes versus paraphrases, but do not specify explicit delimiters or "ignore" instructions to prevent the agent from obeying commands embedded within the transcripts.
Capability inventory: The skill is capable of saving generated content to external notes applications or connectors (e.g., the Scratchpad connector) when available.
Sanitization: There is no mention of sanitizing, escaping, or filtering the input transcripts to remove potentially malicious instructions before they are processed by the agent.

interview-notes