The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted natural language instructions from the user and interpolates them into a generated Python file's instruction dictionary. This creates an attack surface for indirect prompt injection if the generated files are subsequently processed by other automated agents or scoring systems.
Ingestion points: User-provided Task instruction collected during the generation workflow in SKILL.md.
Boundary markers: Absent; instructions are placed directly into a Python string literal within a dataclass configuration.
Capability inventory: The skill performs file-write operations to save the generated task configuration to a user-confirmed path.
Sanitization: No explicit sanitization, escaping, or validation of the user-provided instruction text is mentioned in the instructions.
[COMMAND_EXECUTION]: The skill provides the user with several shell commands (e.g., uv run pytest, python examples/run_empty.py) to validate and test the generated tasks. While these are intended for manual execution by the user in a Docker environment, providing pre-written shell commands is an common pattern that requires user oversight.
Evidence: Verification commands listed in the 'After Generating the Task File' section of SKILL.md.

robolab-taskgen