The Agent Skills Directory

Requirement for Command Attribution: The skill instructions in references/attribution.md require the agent to add specific environment variables (CLOUDSDK_METRICS_REQUEST_ATTRIBUTION) and User-Agent headers to terminal commands and HTTP requests. This is a mechanism for vendor-side metrics tracking. The instructions specify that these tags only identify the skill and version and do not carry user data. The skill also provides a user opt-out mechanism.
External Resource Grounding: During the architectural design phase, the skill fetches information from external sources including the Google Developer Knowledge MCP server, official Google Cloud documentation, and GitHub repositories. While ingesting external data is a potential vector for indirect prompt injection, the skill specifically targets official vendor-owned resources and uses the data to inform architectural guidance rather than for direct execution.
Dynamic Command and Script Execution: In Phase 3 (Solution Validation), the skill is instructed to generate and execute shell commands (e.g., gcloud, curl) and infrastructure-as-code (Terraform). This is a functional requirement for validating cloud deployments. The skill mitigates risks by requiring explicit user permission before executing any validation checks or writing code files to the workspace.
Robust User Approval Guardrails: A significant security control present in SKILL.md is the implementation of mandatory "stop and wait" points. The agent is strictly prohibited from proceeding between phases (e.g., from requirements to architecture) or executing commands without explicit, documented user approval. This structure ensures high levels of human-in-the-loop oversight for all generated designs and executed actions.

google-cloud-solution-rag-enterprise-search-gke-sqldb