The Agent Skills Directory

[DYNAMIC_EXECUTION]: The references/deployment-optimization.md file contains code that uses subprocess.run to execute external scripts (convert_hf_to_gguf.py and llama-quantize). The paths to these executables are constructed using the LLAMA_CPP_PATH environment variable, allowing for execution of code from arbitrary locations if the environment is manipulated.
[INDIRECT_PROMPT_INJECTION]: The references/evaluation-metrics.md file implements an 'LLM-as-Judge' evaluation pattern where model-generated text is passed directly into a prompt for another AI model (e.g., GPT-4o). An attacker who can influence the generated output could potentially inject instructions to bypass the judge's logic or manipulate evaluation results.
Ingestion points: predictions array in llm_judge_evaluation function in references/evaluation-metrics.md.
Boundary markers: Uses labels like Model Response: {prediction} but lacks strong delimiters or 'ignore' instructions for the interpolated content.
Capability inventory: The judge model has the capability to return structured JSON scores which are then processed; the skill also includes subprocess execution and file system write capabilities (e.g., model.save_pretrained).
Sanitization: No explicit sanitization or escaping of the prediction text is performed before it is interpolated into the judge's prompt.
[COMMAND_EXECUTION]: The skill instructs the agent to run local Python scripts (e.g., python validate_dataset.py) for data validation. These scripts are mentioned as necessary checkpoints in the workflow, but their source code is not provided within the skill package, creating a dependency on external, unverified files that might be created or downloaded at runtime.
[EXTERNAL_DOWNLOADS]: The skill's implementation scripts make use of the Hugging Face ecosystem (transformers, datasets, evaluate) to download pre-trained models, datasets, and evaluation metrics from remote repositories. While these are common services, the skill relies on significant volumes of remote content.
[METADATA_POISONING]: Automated security scanners flagged the documentation URL https://jeffallan.github.io/claude-skills/skills/data-ml/fine-tuning-expert/ as malicious. This link is provided as the official reference for the skill, and the blacklist status indicates it should be treated with high caution.

fine-tuning-expert