The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages files and directories using local bash scripts. To prevent shell injection or directory traversal, user-provided inputs (such as experiment names or finding titles) are passed through a slugification function that filters out all characters except lowercase alphanumerics and hyphens.
[PROMPT_INJECTION]: The skill provides strong instructional guidance to the agent to use experiment-specific storage files rather than standard platform defaults (like ephemeral plan files). This behavioral constraint is a functional requirement of the skill's data persistence model and does not constitute a malicious bypass attempt.
[PROMPT_INJECTION]: An indirect prompt injection surface exists where the skill reads content from plans and checkpoints back into the agent's active memory (e.g., via the resume command). This surface is appropriately managed using <file-content> boundary markers which clearly delineate historical data from current instructions.
[EXTERNAL_DOWNLOADS]: The skill utilizes Python 3 for configuration parsing. It assumes the presence of the PyYAML library, which is a standard utility for YAML processing.

doc