How to use the GKE Manifest Generation agent skill

What the skill makes the agent do

Once installed, the agent turns requests such as a described service or a change to application code into Kubernetes manifests for GKE Autopilot and GKE Standard clusters. It writes Deployment, Service and Gateway YAML, and it follows a fixed set of default rules unless you ask otherwise.

• Explicit namespaces on every resource, mapped to the namespace in your active SETTINGS.md
• A dedicated ServiceAccount per microservice instead of the namespace default
• CPU and memory requests and limits on every container
• Non-root containers with privilege escalation disabled and a RuntimeDefault seccomp profile
• Liveness and readiness probes on every Deployment container, plus a startupProbe for slow-starting apps
• Internal ClusterIP services unless public exposure is requested
• Spot VM targeting for dev, test and staging namespaces
• Gateway API routes in preference to legacy Ingress, and inference examples for vLLM, TGI and Gemma

Install the skill

Download the skill ZIP from this page and unzip it. Then place the resulting skill folder in one of these locations, depending on how you want to use it:

• Personal use across projects: unzip into ~/.claude/skills/
• Single project: unzip into the project's .claude/skills/ folder
• Claude apps: upload the ZIP under Settings → Capabilities → Skills

Trigger the skill with natural requests

You do not need to name the skill. The skill description covers requests about creating or modifying GKE deployment manifests, container security contexts, CPU and memory limits, probes, secrets and volumes, Gateway API routes, Spot VMs, and AI inference workloads. Describe the workload and the outcome you want, and mention GKE or Kubernetes so the agent knows the target.

Requests that will not match this skill include live cluster operations and pod troubleshooting, which belong to gke-workload-troubleshooting, and cluster provisioning, which belongs to gke-cluster-creation.

• Write a Deployment and ClusterIP Service for a Go API listening on port 8080 in the payments-dev namespace.
• Add a Gateway and HTTPRoute that exposes my web frontend publicly through GKE Gateway API.
• Create a manifest for a Gemma 2 27B model server with a startup probe and GPU requests.
• Update this Deployment to read its database password from a mounted Secret instead of an environment variable.

Review what the agent produces

The defaults are sensible starting points, but the skill itself says generated YAML should be reviewed against your own cluster policies because the defaults may not match your environment. Check the namespace, ServiceAccount and resource values first. On Autopilot, note that requests and limits must be equal, because Autopilot raises requests to match limits, which can increase cost.

Spot VM targeting adds a nodeSelector and toleration for cloud.google.com/gke-spot. On GKE Standard this assumes a Spot node pool is already configured.

• Confirm the namespace matches your SETTINGS.md configuration.
• Confirm the ServiceAccount is dedicated and not default.
• On Autopilot, confirm requests equal limits.
• Confirm no LoadBalancer or NodePort service was added unless you asked for public access.
• Apply to a test cluster before production.

Safety notes: scripts and network

According to the skill listing, this skill runs no scripts and uses no network access. Generating manifests does not require a cluster. Applying them does, and that step needs Kubernetes tooling and a GKE cluster.

If the agent runs gcloud, REST or MCP commands, the skill's attribution reference requires a metrics environment variable on each gcloud command, and an attribution User-Agent header on direct HTTP or MCP requests. The attribution tag identifies only the skill and its version and carries no user data. The reference also says not to use gcloud config set for this, so the label does not persist into unrelated work.

Tips for better results

Tell the agent the cluster mode (Autopilot or Standard) and the namespace name, since several defaults depend on both. For secrets, say whether the app supports only environment variables; otherwise the skill prefers read-only mounted files. For databases, say explicitly if you need high IOPS or low latency, because the skill reserves SSD storage classes for that case.

• State the cluster mode and namespace in the first message.
• Ask for public exposure explicitly if you want it; otherwise services stay internal.
• For AI model servers, name the model and accelerator so the vLLM, TGI or Gemma reference applies.
• Ask for the startupProbe explicitly for slow-starting apps if you want a specific failureThreshold.