The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches several external assets at runtime. It downloads the A2UI v0.9 composite catalog from Google's gstatic domain. It also fetches typography fonts for CJK and other international scripts directly from the Google Fonts public GitHub repository to ensure proper rendering in generated documents and videos.
[REMOTE_CODE_EXECUTION]: During the deployment process, the skill downloads and executes the uv package manager installer from astral.sh using a piped-to-shell pattern. It also downloads the Google Workspace CLI binary from the official googleworkspace GitHub repository. Additionally, it warm-ups the managed agent sandbox by running pip and npm installs. These actions are performed to bootstrap the required toolchains for the demo environment.
[COMMAND_EXECUTION]: The skill extensively uses subprocess execution to orchestrate infrastructure on Google Cloud. This includes using gcloud to enable APIs, deploy Cloud Run services, manage Secret Manager, and configure Cloud Tasks. It also utilizes the bq CLI for BigQuery dataset management and table loading.
[PRIVILEGE_ESCALATION]: The skill contains logic in its font provisioning script to attempt installation of system font packages using sudo if the current user has the necessary permissions. It also programmatically manages IAM policy bindings to ensure that the Cloud Run service account has sufficient permissions (such as roles/run.invoker or roles/cloudtrace.agent) to operate the demo components. These escalations are scoped to the required setup of the environment.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic Python imports during its pre-deployment preflight check and dependency smoke tests. It uses import() and importlib.import_module() to verify that required first-party and third-party modules are available and compatible before initiating the cloud build process.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data as it performs deep research on customer domains via search_web and processes the findings to synthesize realistic datasets. While this is an inherent part of its functionality, the skill incorporates instructions to handle technical names (Technical Names Isolation) and provides a multi-layer verification engine (verify_and_heal.py) to validate the integrity of the generated environment.

ge-demo-generator