The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill utilizes the !command syntax in SKILL.md to run neuron-ls | head -3 during load time. This is used for hardware platform detection and is a standard, benign use of dynamic context in a developer environment for the purpose of identifying the available Neuron devices.
[COMMAND_EXECUTION]: The instructions provide templates and workflows for the agent to execute shell commands such as neuron-ls, source, python, and neuron-top, and to run Python scripts that invoke the Neuron compiler (neuronx-cc). These capabilities are necessary and expected for the primary purpose of debugging, testing, and profiling NKI kernels on specific hardware.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface where it processes user-provided NKI kernels and test scripts to facilitate debugging.
Ingestion points: User-provided kernel code and external kernel files identified via the argument hint in SKILL.md.
Boundary markers: The instructions do not explicitly require the agent to wrap user-provided code in specific delimiters or safety warnings before execution.
Capability inventory: The skill leverages subprocess execution (via python), environment variable modification (via os.environ), and file system writes (for profiling artifacts in ./output).
Sanitization: There is no mention of sanitizing or validating the contents of user-supplied scripts or kernel files before they are executed in the local environment.

neuron-nki-debugging