The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user to download and execute an installation script for the uv package manager using curl -LsSf https://astral.sh/uv/install.sh | sh. This pattern involves piping remote content directly to a shell. The source astral.sh is the official domain for Astral, a well-known provider of high-performance Python developer tools.
[EXTERNAL_DOWNLOADS]: The skill clones the nanochat source code repository from GitHub (https://github.com/karpathy/nanochat.git) to obtain the training harness and scripts.
[COMMAND_EXECUTION]: Multiple commands are provided to execute local Python scripts and shell scripts for training and evaluation, including torchrun for distributed training and subprocess.run within Python examples to manage training processes.
[DATA_EXFILTRATION]: The instructions describe serving a web-based chat interface (scripts.chat_web) and suggest accessing it via the host's public IP on port 8000. This involves opening a network port to external traffic, which is a standard requirement for the described chat UI functionality.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of a chat environment that ingests untrusted user input.
Ingestion points: Data enters the system through the scripts/chat_web.py and scripts/chat_cli.py interfaces during inference.
Boundary markers: None are explicitly defined in the skill instructions to separate user input from system instructions within the generated chat environment.
Capability inventory: The environment includes capabilities for file system writes (checkpointing), network serving (web UI), and execution of training processes via torchrun and subprocess.
Sanitization: No specific sanitization or filtering logic for external prompt content is detailed in the provided instructions.

nanochat-llm-training