Kubernetes Manifest Author

Kubernetes Manifest Author

Overview

Produce correct, secure, and reliable Kubernetes YAML for deploying stateless and stateful workloads. This skill turns an application description into a complete manifest set: workload (Deployment/StatefulSet/CronJob), Service, optional Ingress, probes, resource governance, security context hardening, and configuration wiring.

Keywords: kubernetes, k8s, manifest, yaml, deployment, statefulset, service, ingress, liveness probe, readiness probe, startup probe, resources requests limits, securityContext, runAsNonRoot, podSecurityStandards, HPA, autoscaling, ConfigMap, Secret, PodDisruptionBudget, rollout, kubectl, helm, kustomize.

This skill is opinionated toward the restricted Pod Security Standard and production SRE defaults. Loosen only when the user has a concrete reason.

Workflow

1. Gather the workload shape. Determine: container image + tag (never :latest), listening port(s), stateless vs stateful, replica count, environment/config inputs, secrets, and exposure (cluster-internal, Ingress, LoadBalancer). If the user omits something, apply the safe defaults in references/best-practices.md and state the assumption.

1. Pick the workload kind using the decision table below. Most web/API services are Deployment.

1. Author the workload manifest. Start from templates/deployment.yaml. Always fill in: explicit labels, image with pinned tag, resource requests AND limits, all three probe types where appropriate, a hardened securityContext, and a RollingUpdate strategy.

1. Add the Service. Use templates/service.yaml. Match selector to the Pod template labels exactly. Default to ClusterIP; use LoadBalancer/NodePort only when externally requested.

1. Add exposure and resilience extras as needed: Ingress, HorizontalPodAutoscaler, PodDisruptionBudget, ConfigMap/Secret. See references/resource-catalog.md.

1. Wire configuration. Inject config via envFrom/env (ConfigMap) and secrets via secretKeyRef or mounted volumes. Never hardcode secrets in the manifest.

1. Validate. Run scripts/validate_manifest.py to catch the most common production mistakes (missing limits, :latest, root user, no probes, mismatched selectors). If kubectl is available, also run kubectl apply --dry-run=server and kubectl explain for field checks.

1. Explain the manifest to the user: label the security/reliability choices you made and what they should tune (resource numbers, replica count, probe paths).

Decision Framework — which workload kind?

| Situation | Kind | |-----------|------| | Stateless HTTP/gRPC API or web app | Deployment | | Needs stable network identity / ordered start / per-pod storage (DBs, Kafka) | StatefulSet | | One pod per node (log/metrics agents, CNI) | DaemonSet | | Run-to-completion task | Job | | Scheduled run-to-completion task | CronJob |

Probe decision framework

Probes are the most common source of production incidents. Apply this:

• readinessProbe — ALWAYS for anything in a Service. Gates traffic; failing readiness removes the pod from endpoints without restarting it. Should check that dependencies (DB, cache) the pod needs to serve are reachable.
• livenessProbe — only when the app can deadlock in a way a restart fixes. Make it CHEAP and dependency-free (it should NOT check the database — a DB outage shouldn't restart every pod). A bad liveness probe causes restart storms.
• startupProbe — for slow-starting apps (JVM, migrations). Disables liveness/readiness until the app has booted, so you can keep liveness aggressive without killing slow boots.

Probe handler choice: prefer httpGet for HTTP servers, grpc for gRPC, exec only as a last resort (forks a process each check).

See worked numbers and timing math in references/best-practices.md.

Security context checklist (restricted PSS)

Every production Pod should set, at the container level unless noted:

securityContext:
  runAsNonRoot: true
  runAsUser: 10001            # any non-zero UID
  allowPrivilegeEscalation: false
  readOnlyRootFilesystem: true
  capabilities:
    drop: ["ALL"]
  seccompProfile:             # pod-level also acceptable
    type: RuntimeDefault

readOnlyRootFilesystem: true requires mounting emptyDir volumes for any writable paths (e.g. /tmp, cache dirs). The template shows this pattern.

Worked example

See examples/web-api.md for a full request → manifest walkthrough: a Node.js API with config, secrets, an HPA, a PDB, and an Ingress — including the reasoning for every non-default field.

Best Practices

• Pin image tags (or digests); set imagePullPolicy: IfNotPresent for pinned tags.
• Set resource requests (for scheduling/QoS) AND limits (to cap blast radius). Set CPU request = limit only when you need predictable latency; otherwise omit the CPU limit to avoid throttling (see reference).
• Use the recommended common labels (app.kubernetes.io/*) so Services, HPAs, and dashboards select consistently.
• Minimum 2 replicas + a PodDisruptionBudget for any service that must stay up during node drains.
• Keep liveness probes dependency-free; put dependency checks in readiness.
• Prefer RollingUpdate with maxUnavailable: 0 for zero-downtime when you have spare capacity.
• One concern per manifest file; order resources so dependencies (ConfigMap, Secret) come before consumers, or use separate files applied together.
• Run scripts/validate_manifest.py before shipping.

Common Pitfalls

• :latest tags — non-reproducible rollouts and no rollback guarantee.
• No resource limits — one pod can starve a node; pods get BestEffort QoS and are evicted first.
• Liveness probe checking the database — turns a dependency outage into a cluster-wide restart storm.
• Selector/label mismatch — Service routes to nothing; kubectl get endpoints is empty.
• Running as root — blocked by restricted PSS; an escape is far more dangerous.
• readOnlyRootFilesystem: true without writable emptyDir mounts — app crashes trying to write /tmp.
• Aggressive initialDelaySeconds on liveness for slow apps — use a startupProbe instead.
• Single replica + maxUnavailable: 1 — guaranteed downtime on every rollout and node drain.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JayRHa
• Source: JayRHa/AgentSkills
• License: MIT

Install and usage instructions live in the source repository linked above.