The Agent Skills Directory

[SAFE]: The skill consists of instructional content and structural patterns for AI evaluation loops. It does not contain executable code or malicious tools.
[PROMPT_INJECTION]: The skill describes patterns for ingesting external data (tasks, outputs) into refinement loops (Indirect Prompt Injection surface). Ingestion points: 'task', 'output', and 'criteria' variables in the Python snippets in SKILL.md. Boundary markers: Absent from snippets. Capability inventory: Conceptual 'llm()' and 'run_tests()' functions. Sanitization: Absent from template code.
[COMMAND_EXECUTION]: The 'Code-Specific Reflection' pattern illustrates the process of verifying generated code with a test runner. This is documented as a developmental workflow pattern and is not an active execution risk within the skill.

agentic-eval