The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted interview transcript files (.txt, .docx, .srt, .vtt). To mitigate the risk of indirect prompt injection from respondent text, the skill includes a built-in detection mechanism in scripts/batch_prepare.py and scripts/number_lines.py. These scripts use comprehensive regular expression patterns to identify and flag instructions that might attempt to hijack the model's behavior (e.g., 'ignore previous instructions', 'system prompt', etc.). The documentation also explicitly warns against executing text found inside transcripts.
[DYNAMIC_EXECUTION]: The script scripts/coverage_gaps.py uses importlib.util to dynamically load a sibling script (verify_quotes.py). This is performed within the local execution environment and is used to ensure consistency in text normalization logic across different modules without duplicating code.
[COMMAND_EXECUTION]: The script scripts/calibrate_threshold.py executes another local skill script via subprocess.run. The command is constructed using a fixed list of arguments and the local Python executable (sys.executable), ensuring the execution remains within the skill's context and is not susceptible to arbitrary command injection.
[SAFE]: The skill implements protective measures against 'Zip Bomb' and 'Billion Laughs' XML expansion attacks in scripts/batch_prepare.py and scripts/number_lines.py. It uses the expat parser to explicitly reject Document Type Definitions (DTDs) and enforces a maximum file size limit when unpacking .docx files.

interview-mapper