The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, specifically interview transcripts and notes provided by the user (Phase A: Intake).
Ingestion points: Raw material (transcripts, notes, memory dumps) entered by the user.
Boundary markers: Absent; there are no explicit instructions for the agent to ignore potentially malicious commands embedded within the transcript data.
Capability inventory: Writing structured markdown files to an interviews/ directory.
Sanitization: Absent; the skill prioritizes verbatim extraction of customer phrases to preserve vocabulary data.
[DATA_EXPOSURE]: The skill accesses local files (e.g., QUESTIONS.md, HYPOTHESES.md) to map interview answers to specific questions. It writes debrief outputs to a local interviews/ directory. These operations are scoped to the specific project and do not involve sensitive system directories, environment variables, or hardcoded credentials.
[SAFE]: The skill maintains a legitimate posture by including clear refusal conditions. It explicitly declines to synthesize data across multiple interviews, manufacture answers, or process role-played/AI-generated content, ensuring it only handles real evidence as intended. No obfuscation, remote dependencies, or persistence mechanisms were detected.

asb-interview-debrief