The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands to build and install operator artifacts.
Evidence: Phase 3.4 includes shell execution for source ${ASCEND_HOME_PATH}/set_env.sh, bash build.sh, and pip install output/ascend_kernel*.whl --force-reinstall --no-deps.
Context: These are standard operations within the Ascend development environment for compiling and testing NPU kernels.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection (Category 8).
Ingestion points: The skill reads operator design documents (design.md) and C++ source files (op_host/*.cpp, op_kernel/*.cpp) to learn the operator's logic (SKILL.md, Section 1.1).
Boundary markers: Absent. There are no explicit instructions or delimiters used to separate the content of these processed files from the agent's instructions.
Capability inventory: The skill has the capability to modify local files, execute shell scripts (build.sh), and perform package installations (pip install).
Sanitization: Absent. The content of the read files is not sanitized or validated before being processed in the context of the optimization task.

ascendc-operator-performance-optim