The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it processes external session captures and walkthrough transcripts to generate documentation. \n
Ingestion points: Untrusted data enters the agent context through the --input arguments of the execution scripts and the --context parameters in the prompt templates (e.g., requirements-doc-agent.md). \n
Boundary markers: The skill uses template headers and structured Markdown sections but lacks robust delimiters or sanitization for the ingested text. \n
Capability inventory: The agent has access to Read, Write, and Bash tools, which are used to generate and modify documentation in the exploration/ directory. \n
Sanitization: No explicit sanitization or escaping of the source text is performed before processing. \n
Context: This is a standard functional risk for documentation agents and is mitigated by explicit instructions to mark inferred content with [UNCONFIRMED] and use [NEEDS HUMAN INPUT] for gaps.\n- [COMMAND_EXECUTION]: The skill provides Bash command examples for running local Python scripts to scaffold requirements documents and verify gap thresholds. These scripts (business_requirements_capture_execute.py, etc.) use standard libraries and do not perform network requests or execute untrusted code.

capturing-business-requirements