The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the ingestion of large datasets for model training and fine-tuning, which presents a surface for indirect prompt injection if the source data is untrusted.
Ingestion points: Training data enters the agent context via the trainset variable in SKILL.md and by loading external files such as reviews_labeled.json and support_messages_labeled.json in examples.md.
Boundary markers: The provided code snippets do not demonstrate the use of explicit delimiters or instructions to the model to ignore embedded commands within the training examples.
Capability inventory: The skill utilizes the dspy library, which performs network operations to interface with LLM provider APIs (OpenAI, Together AI) and includes functionality to write and read local JSON files for model configuration (.save() and .load()).
Sanitization: No explicit validation or sanitization of the input data strings is present in the examples provided.

dspy-bootstrap-finetune