The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes system commands using the python -m skore_skills utility to manage project status, styling, notebook filling, and git operations.
Evidence includes calls to status, eda stamp, style, notebook fill, site build, and git end-turn across various files like SKILL.md and references/extra_analyses.md.
[DYNAMIC_EXECUTION]: The skill dynamically assembles and executes a Python script (materialize.py) at runtime to perform data profiling.
Evidence: The procedure in SKILL.md (Step 2) and templates/materialize.py instructs the agent to copy a template, perform string substitutions (e.g., <LOAD_RAW_DATA>, <TARGETS>), and execute the resulting script using the local Python interpreter.
[INDIRECT_PROMPT_INJECTION]: The skill has a defined attack surface for indirect prompt injection because it ingests untrusted user data and uses findings derived from that data as context for subsequent agent reasoning.
Ingestion points: The materialize.py script reads user-provided raw data files. The agent then reads the generated extras.json and data_analysis.md back into its context.
Boundary markers: The instructions do not specify the use of clear delimiters or "ignore instructions" warnings when interpolating data-derived summaries into prompts for other tools, such as the research-ml-practice skill.
Capability inventory: The agent has the capability to execute shell commands, write to the filesystem, and load other functional skills based on analysis results.
Sanitization: No explicit sanitization or escaping of data values (e.g., column names, category labels, or data summaries) is performed before they are used in natural language reasoning steps.

explore-ml-data