The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content (PDF, PPT, Excel, Word) and outputs it as structured Markdown to be used by downstream agents. If the downstream agent treats this parsed data as instructions, it is vulnerable to indirect prompt injection.
Ingestion points: Processes files via the file_path parameter in SKILL.md and scripts/run_doc_parse.py.
Boundary markers: None explicitly defined in the skill instructions to separate document content from agent instructions.
Capability inventory: The skill possesses file system read and write capabilities (writing to examples/snse_survey/skill_outputs/doc_parse.json).
Sanitization: The skill relies on an external security-scan mentioned in the documentation, but does not implement internal sanitization of the parsed Markdown text.
[COMMAND_EXECUTION]: The script scripts/run_doc_parse.py dynamically modifies the Python search path using sys.path.insert to include the core directory relative to its location. This is a common pattern for local project imports but ensures that code in the core directory is executed when AgentSession is imported.
[DATA_EXFILTRATION]: The skill reads data from local paths and writes the parsed output to a JSON file in the examples/snse_survey/skill_outputs/ directory. While this is the intended purpose, it represents a local data movement pattern.

antinet-doc-parse