The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external acceptance criteria from YAML files through an LLM to generate code, creating a potential surface for indirect prompt injection.
Ingestion points: Acceptance criteria are read from files like input/cart-promo.yaml and passed to the LLM in scripts/ai-gen.py.
Boundary markers: The skill uses a fixed system prompt to constrain LLM behavior, though it does not use explicit delimiters for the input data.
Capability inventory: The skill writes generated code to the filesystem and executes syntax validation commands.
Sanitization: Includes a heuristic-based confidence scoring rubric and a mandatory human-in-the-loop review process before merging.
[COMMAND_EXECUTION]: The skill executes shell commands to perform syntax and compilation checks on generated test files.
Evidence: npx tsc --noEmit, python -m py_compile, and node --check are used in the validation step defined in SKILL.md.
[DYNAMIC_EXECUTION]: The skill generates source code files and dynamically checks for the existence of modules at runtime.
Evidence: The save_test function in references/generation-and-scoring.md writes string content from the LLM to .js or .py files.
Evidence: The module_exists helper function uses importlib.util.find_spec to check if modules imported in the generated code are present in the environment.

ai-test-generator