The Agent Skills Directory

[SAFE]: The skill is primarily a procedural guide and toolset for interacting with official Google Cloud Platform (GCP) services (Vertex AI, GCS) and Hugging Face. The scripts provided are utilitarian in nature, focusing on data formatting, cost estimation, and job management.
[EXTERNAL_DOWNLOADS]: The skill downloads configuration and datasets from well-known and trusted services, including official Google Cloud APIs and the Hugging Face public registry. Dependencies in references/requirements.txt are restricted to official, version-pinned libraries (e.g., google-cloud-aiplatform, google-genai).
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from local files, GCS buckets, and Hugging Face repositories, which presents a common attack surface for indirect prompt injection.
Ingestion points: scripts/prepare_dataset.py and scripts/calculate_cost.py read data from user-specified or remote sources.
Boundary markers: The skill explicitly instructs the agent to present data samples and column mapping recommendations to the user for confirmation before any data processing occurs.
Capability inventory: The skill can execute shell commands (gcloud), run Python scripts, and perform network operations to GCP endpoints.
Sanitization: The skill employs structural validation of JSONL schemas and requires mandatory interactive human checkpoints (Tier M) before any state-changing operations are triggered, effectively mitigating unsupervised obedience to embedded data instructions.
[SAFE]: The skill instructions emphasize defensive practices, such as verifying the existence of datasets before submission, prohibiting silent retries, and requiring explicit user permission before searching buckets or installing components.

agent-platform-tuning