The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's documentation and Python scripts (e.g., reference/sft/sales_extractor_training.py and reference/deployment.md) reference the installation of Ollama and Rust using standard curl | sh patterns. These are well-known services and the commands are provided for user-initiated installation or as informational print statements.
[EXTERNAL_DOWNLOADS]: The skill downloads numerous Python dependencies (such as unsloth, trl, vllm, and torchao) from official package registries (PyPI) and reputable GitHub organizations like unslothai. These are standard requirements for its stated purpose of model training.
[COMMAND_EXECUTION]: The deployment documentation (reference/deployment.md) provides Docker commands that interact with the host system by mounting local directories and reading the user's public SSH key (~/.ssh/id_rsa.pub) to facilitate secure container access. These are standard development practices.
[PERSISTENCE]: The skill implements a local analytics sidecar by instructing the agent to write session outcomes to ~/.claude/skill-analytics/last-outcome-unsloth-training-skill.json. This is used for internal tracking of training success and does not exhibit malicious behavior.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (sales call transcripts) for training purposes. It uses regular expressions to extract data from specific XML-like tags (e.g., <reasoning>, <answer>). While this is a common attack surface for indirect injection, the risk is mitigated by the specific extraction logic and the intended use case of processing data into structured JSON formats.

unsloth-training-skill