The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local Python scripts (search_evidence.py, prepare_pdftotext.py, create_specialized_skill.py) to perform its primary tasks. The agent is instructed to run these scripts to retrieve evidence and generate new specialized skills. This is the intended functional design of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided local files (Markdown, text). There is a potential risk that these files could contain adversarial instructions designed to influence the agent's behavior during the retrieval and answering process.
Ingestion points: The scripts/search_evidence.py script recursively scans directories and reads file contents into the agent's context.
Boundary markers: The skill instructions implement a robust evidence contract, requiring the agent to classify statements as Direct, Synthesis, or Inference and map them to specific source IDs, which serves as a mitigation against following embedded instructions.
Capability inventory: The skill allows the agent to execute local Python scripts and write to the file system (specifically when creating new specialized skills).
Sanitization: Content is normalized for search but not sanitized for embedded instructions.
[PRIVILEGE_ESCALATION]: The scripts/create_specialized_skill.py script utilizes chmod 0o755 to ensure the retrieval script in newly generated specialized skills is executable. This is a standard setup procedure for the creation of functional tools within the workspace.

z-grounded-source-qa