The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches pre-trained model weights and configurations from the Hugging Face Hub (e.g., mlx-community/Llama-3.2-3B-Instruct-4bit). These are downloads from a well-known machine learning service.
[COMMAND_EXECUTION]: Instructs the agent on using the mlx-lm command-line utility for model generation, interactive chat, conversion, fine-tuning, and server hosting.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data for inference and fine-tuning, which is a standard surface for LLM tooling.
Ingestion points: Processes user prompts via the generate function and training data from local JSONL files (e.g., train.jsonl) located in the ./data directory.
Boundary markers: Recommends the use of tokenizer.apply_chat_template() as a best practice to define role boundaries and prevent instruction leakage.
Capability inventory: Includes capabilities for text generation, model conversion, weight quantization, and LoRA adapter training.
Sanitization: Relies on standard framework-provided tokenization and chat templates to structure inputs for instruction-tuned models.

mlx