The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes project-specific data such as problem framing translations, status reports, and dataset metadata to generate machine learning code. This creates a potential surface for indirect prompt injection if the local project environment or metadata is maliciously modified.
Ingestion points: Data entering the agent context includes output from skore_skills status and skore_skills frame show (recorded in SKILL.md and references/source-binding.md).
Boundary markers: The skill implements a 'Pre-flight' checklist for manual verification but does not explicitly instruction the agent to sanitize external content before interpolation.
Capability inventory: The skill has extensive file-writing capabilities (creating pipeline.py, experiment files, and tests) and can execute these scripts via a specialized smoke-test command.
Sanitization: No explicit validation or escaping of external metadata is described before it is written into source files.
[COMMAND_EXECUTION]: Orchestrates the ML lifecycle by executing numerous subcommands of the skore_skills CLI tool, including status monitoring, design validation, and running smoke tests.
[DYNAMIC_EXECUTION]: Performs automated generation and modification of Python scripts for ML pipelines and experiments. These generated files are then executed in a controlled environment to verify functionality.

build-ml-pipeline