The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures untrusted user input and interpolates it into prompts for the 'Explore' and 'interviewer' subagents.
Ingestion points: Stage 1.1 extracts the core problem from user input, and Stage 2.2 captures the user's answers to interview questions.
Boundary markers: Absent. The instructions do not define delimiters (e.g., XML tags or triple quotes) or provide 'ignore embedded instructions' warnings when passing user data to subagents.
Capability inventory: The skill has access to Write, Agent, Read, Grep, and Glob tools, creating a functional surface if a subagent is successfully manipulated by user input.
Sanitization: Absent. There is no evidence of input validation, escaping, or filtering for strings provided by the user before they are processed by the subagent.
[COMMAND_EXECUTION]: The skill uses the Write tool to save insights to a file path constructed from a user-supplied topic.
Evidence: In Stage 3.3, the skill performs a Write("deep-interview-outputs/[topic-slug]/insights.md", insights_content) operation. The topic-slug is derived from the user's initial input. Without proper sanitization of this slug, there is a theoretical risk of path traversal if the underlying platform's Write tool does not enforce strict directory boundaries.

deep-interview