How to use the GKE AI Workload Migration Guide agent skill

What is an agent skill and what does this one do?

An agent skill is a folder with a SKILL.md file that tells an AI agent how to handle a specific kind of task. When the skill is installed and your request matches its description, the agent follows the skill's workflow instead of improvising.

This skill covers migrations of existing AI inference workloads to self-hosted inference on GKE. The agent works through four phases: discovery, design, implementation and validation. In discovery it reads your existing configuration with read-only gcloud list and describe commands. It then writes Kubernetes manifests to your current directory and gives you the gcloud and kubectl commands to run yourself. The default serving setup uses vLLM, and the service is exposed through the GKE Gateway API with a regional internal Application Load Balancer.

• Reads existing configurations with read-only gcloud commands after you grant permission
• Writes Kubernetes manifests to disk before any change is applied
• Provides gcloud and kubectl commands for you to review and run
• Uses Custom Compute Classes for accelerator obtainability and GKE managed GPU drivers
• Uses GKE Workload Identity for Google Cloud API access

Install the skill (ZIP download or Claude apps)

Download the ZIP for this skill from this page and unzip it. You can install it in one of two places. For a personal setup available across projects, unzip it into ~/.claude/skills/. For a single project, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Check the requirements before you start

This skill is for people with real infrastructure to move. You need a Google Cloud account with an existing AI inference workload, the gcloud CLI installed and authenticated, the kubectl CLI, and access to a GKE cluster or permission to create one. You also need GPU quota. Cluster and billing changes carry real cost and outage risk, so do not start without a Google Cloud project and GKE experience.

• Google Cloud account with an existing AI inference workload to migrate
• gcloud CLI installed and authenticated
• kubectl CLI
• Access to a GKE cluster or permission to create one
• GPU quota

Trigger the skill with a natural request

You do not need a special command. Describe your situation in plain language and mention the source platform and the goal. The agent matches the request against the skill description, which covers hardware sizing, model staging, manifest generation, validation and traffic cutover.

If you do not mention MCP, the agent will add one sentence in its first discovery reply noting that an automated alternative exists and you can switch to it at any time. If you say you do not want MCP, it proceeds with the manual workflow.

• "I run a Gemma model on Cloud Run. Help me migrate it to self-hosted inference on GKE."
• "Our Vertex AI endpoint is expensive. Walk me through moving it to GKE with GPUs."
• "No MCP, let's do this manually. Start with discovery on my existing Cloud Run service."

Understand the safety rules: scripts and network access

The skill is marked as running scripts and using the network. Its rules say that in discovery the agent may run read-only gcloud inspection commands, but in design and implementation it writes manifests and shows commands instead of running them. Mutating commands such as apply, create, delete and cluster or IAM changes run only if you explicitly ask the agent to run them, and then the agent reports each command's actual output.

The skill also requires a usage attribution step. Before any gcloud, REST or MCP command, the agent must follow references/attribution.md, which tags commands with an environment variable set inline for that command only. The tag identifies the skill and its version and carries no user data. Gated models that need a Hugging Face token must reference a Kubernetes secret rather than a literal token value, and the agent is instructed to warn you about the security risk.

• Read-only list and describe commands run in discovery after you grant permission
• Mutating commands run only on your explicit request
• Attribution is set inline on each gcloud command, never with gcloud config set
• HF_TOKEN values are referenced with secretKeyRef, never written into manifests

Know when to use a different skill or tool

If you are deploying a new model server on GKE with no existing workload, the skill tells the agent to stop and recommend the gke-inference skill instead. If you want the agent to change cloud resources automatically, the skill hands off to the Gemini Cloud Assist MCP server. In that case the agent stops the manual workflow and points you to the Gemini Cloud Assist MCP documentation and the Intent to Infrastructure codelab.

• New deployment with nothing to migrate: use gke-inference
• Automated changes to cloud resources: use the Gemini Cloud Assist MCP server
• Manual, reviewable migration: this skill