The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands and scripts to manage the project lifecycle, including creating directories (mkdir), copying template files (cp), and setting execution permissions (chmod). It also reads configuration from /etc/Ascend/ascend_cann_install.info to locate the vendor's CANN toolkit. These actions are standard for the target development environment and restricted to relevant project paths.
[EXTERNAL_DOWNLOADS]: The build.sh template includes a mechanism to install the wheel package from the official Python Package Index (PyPI). This is a well-known service, and the installation is a routine step in preparing the build environment.
[INDIRECT_PROMPT_INJECTION]: The skill takes an operator name provided by the user and uses it to generate file paths and C++ source code. To mitigate injection risks, the skill explicitly enforces naming conventions (alphanumeric characters and underscores only) and performs safety checks to ensure existing directories are not overwritten.
Ingestion points: The op_name variable provided by the user in SKILL.md.
Boundary markers: Not explicitly used in command interpolation.
Capability inventory: File system modification (mkdir, cp), script execution (bash), and permission changes (chmod) across the skill and its templates.
Sanitization: Validation logic requiring the operator name to be snake_case and alphanumeric, which prevents path traversal or shell character injection.
[SAFE]: The C++ templates use dynamic library loading (dlopen, dlsym) to interface with NPU-specific libraries like libopapi.so. This is a standard architectural requirement for Ascend operator extensions and does not represent a malicious execution pattern.

ascendc-operator-project-init