The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Fetches and executes the official Ollama installation script (https://ollama.com/install.sh) via a shell pipe. This is a well-known service used for local LLM inference.
[REMOTE_CODE_EXECUTION]: Downloads and executes the official Unsloth auto-installation script (https://raw.githubusercontent.com/unslothai/unsloth/main/unsloth/_auto_install.py) directly through the Python interpreter.
[REMOTE_CODE_EXECUTION]: Installs the Rust toolchain using the official installer script from sh.rustup.rs, which is a well-known service for Rust development.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of sales call voice transcripts to generate training datasets.
Ingestion points: Transcripts are ingested in reference/sft/sales_extractor_training.py and reference/sft/generate_training_data.py.
Boundary markers: No specific delimiters are used to isolate the untrusted transcript content from instructions.
Capability inventory: The skill has the capability to write files to the local system via model.save_lora and model.save_pretrained_merged.
Sanitization: No sanitization or validation of the transcript content is performed before it is processed by the model.
[DATA_EXFILTRATION]: Documentation examples suggest using environment variables to pass sensitive information, such as SSH_KEY for Docker setup and API keys for vLLM or Weights & Biases logging. While these are common practices, users should ensure these secrets are managed securely.

unsloth-training