The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted content including candidate resumes, job descriptions, and interview transcripts to generate automated reports.
Ingestion points: Untrusted data is ingested into the agent context in task-1-pre-interview.md (JD and Resumes), task-2-post-interview-summary.md (Transcripts and Notes), and task-3-post-interview-questions.md (Transcripts).
Boundary markers: The prompt templates utilize standard Markdown headers (e.g., ### 输入上下文) to separate inputs, but they lack robust delimiters or specific instructions for the agent to disregard commands embedded within the data.
Capability inventory: The skill instructs the agent to execute shell commands using the memocli tool, including piping content derived from LLM output (e.g., echo "[content]" | memocli create-entity).
Sanitization: There is no evidence of sanitization or escaping of the ingested data before it is interpolated into prompts or used in shell command strings.
Risk: Malicious text within a transcript or resume could be used to manipulate the agent's evaluation results or attempt to escape the echo command structure through the use of shell-sensitive characters like backticks or dollar signs.
[EXTERNAL_DOWNLOADS]: The skill declares a dependency on github/cafe3310/agent-skill-memories-off in its frontmatter. This is a vendor-owned resource from the same author used for local entity management.

interview-processor