The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes experimental data from JOURNAL.md and log files to generate scientific reports.\n
Ingestion points: JOURNAL.md, IGNORED_RUNS.md, and plaintext log files (*.log) in the runs/ directory.\n
Boundary markers: The skill instructions include explicit "no hallucination" and "no editorialization" directives for the agent when synthesizing this data.\n
Capability inventory: The agent has the ability to write files and execute Python scripts via the shell.\n
Sanitization: No explicit sanitization of log content is implemented in the provided regex-based parsing scaffolds beyond standard JSON loading.\n- [DYNAMIC_EXECUTION]: The skill generates and executes Python scripts (train.py, eval.py, plot_curves.py) using provided Typer-based scaffolds.\n
Execution method: Scripts are executed in managed environments using uv run or pixi run.\n
Context: The scripts are artifacts of the experimentation process and are generated from known templates to ensure a canonical directory structure.\n- [COMMAND_EXECUTION]: The skill uses shell commands to manage virtual environments and run experiments.\n
Functionality: Shell access is used to invoke uv and pixi for running the generated experimentation scripts.\n- [EXTERNAL_DOWNLOADS]: The skill references standard machine learning package registries and specific indices for GPU support.\n
Evidence: Mentions the official PyTorch CUDA index for dependency resolution within PEP 723 metadata blocks.

ml-experimentation