The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands to interact with the environment and the data-designer CLI. This includes resolving the executable path, validating configuration scripts, and generating datasets (e.g., data-designer validate, data-designer preview, and data-designer create in workflows/autopilot.md and workflows/interactive.md).
[COMMAND_EXECUTION]: The skill generates Python scripts based on user requirements and subsequently executes them via the data-designer CLI. While this is the primary function of the skill, it involves the runtime execution of dynamically generated code.
[PROMPT_INJECTION]:
Indirect Prompt Injection: The skill is vulnerable to indirect prompt injection through the ingestion of external seed datasets as described in references/seed-datasets.md.
Ingestion points: The skill reads local files (Parquet, CSV, JSON, JSONL) as seed sources to bootstrap data generation.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to distinguish between valid seed data and potential malicious instructions embedded within the datasets.
Capability inventory: The skill possesses the capability to write Python scripts and execute shell commands through the data-designer toolset.
Sanitization: No evidence of sanitization or validation of the content within seed datasets is present before the data is interpolated into Jinja2 templates.
Instruction Override: The Troubleshooting section in SKILL.md contains an instruction for the agent to suggest retrying commands with the sandbox disabled if network errors occur. Although this action is gated by a request for user permission, it guides the agent to bypass platform security constraints.

data-designer