The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes visual content that may include text or instructions capable of influencing the agent's logic or the content of revision requests.
Ingestion points: The score_review.py script reads JSON data produced from visual inspections described in SKILL.md.
Boundary markers: The instructions lack explicit delimiters or markers to isolate untrusted visual data from agent instructions.
Capability inventory: The skill can route RevisionRequest contracts to multiple peer agents (e.g., brand-intelligence, composition-director) and execute a local Python scoring script.
Sanitization: Input validation is limited to type casting JSON values to floats in the scoring logic, with no text-based filtering.
[DYNAMIC_EXECUTION]: The test utility scripts/test_gates.py utilizes dynamic loading to import local modules for verification.
Evidence: The script uses importlib.util.spec_from_file_location to load score_review.py from a path computed via the __file__ attribute. This is a standard testing pattern with no dependencies on external or untrusted inputs.

visual-qa