The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a local Python script (scripts/plan_lint.py) designed to validate the completeness of experiment matrix files. The script performs static analysis using regular expressions and does not execute arbitrary code or perform unsafe system operations.
[EXTERNAL_DOWNLOADS]: The documentation references and provides installation instructions for widely-used open-source research and data science tools, including DVC, MLflow, Weights & Biases, Hydra, Snakemake, and various Python scientific libraries (scikit-learn, statsmodels, PyMC). These references point to official documentation and reputable GitHub repositories.
[DATA_EXFILTRATION]: While the skill involves the use of experiment tracking platforms (MLflow, Weights & Biases), the instructions explicitly recommend using offline modes for sensitive data to prevent unintended data exposure.
[PROMPT_INJECTION]: The skill is designed to process user-provided research ideas and project data, which introduces an inherent surface for indirect prompt injection. However, the skill employs structured templates and specific verification checkpoints that help maintain the agent's focus on the research planning task.
Ingestion points: User research ideas and project-specific files (e.g., PROJECT_PLAN.md, experiment_matrix.md).
Boundary markers: The skill does not currently use explicit delimiters to isolate untrusted input.
Capability inventory: The agent has access to the provided plan_lint.py script and is instructed on using command-line tools like Hydra and Snakemake.
Sanitization: No specific input sanitization or validation of the content of user-provided ideas is implemented beyond the linter's structural checks.

light-research-plan