菜单
Use when reviewing Kubernetes manifests for cost optimization, resource right-sizing, or identifying over-provisioned workloads. Only triggers when K8s manifests are detected.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | kubernetes-cost-review |
| description | Use when reviewing Kubernetes manifests for cost optimization, resource right-sizing, or identifying over-provisioned workloads. Only triggers when K8s manifests are detected. |
Cost analysis framework for Kubernetes workloads. Reviews resource requests and limits, scaling configuration, storage provisioning, and environment-specific sizing. Every workload should be right-sized for its actual needs, with scaling policies that prevent both waste and outages.
Only triggers when there is clear evidence of Kubernetes usage. When ambiguous, ask the user before proceeding.
terraform-cost-review insteadScan all K8s manifests and inventory every workload. Categorize by:
See sizing-reference.md for workload sizing guidelines.
For each workload, check:
For each workload, recommend resource values for dev, staging, and production. See sizing-reference.md for common sizing patterns.
If deprecated APIs with cost implications are found (e.g., deprecated autoscaling/v2beta1 HPA that affects scaling behavior, deprecated policy/v1beta1 PDB), flag them in findings and recommend running deprecation-check-and-upgrade for full migration guidance.
| Trap | Typical Impact | Fix |
|---|---|---|
| No resource requests set | Scheduler can't bin-pack, nodes under-utilized | Set requests based on actual usage patterns |
| Requests == Limits for all containers | No burst headroom, over-provisioned for steady state | Set requests to P50, limits to P99 of actual usage |
| No HPA on variable-load workloads | Paying for peak capacity 24/7 | Add HPA with CPU/memory or custom metrics |
| LoadBalancer Service per microservice | Cloud LB costs $15-25/month each | Use Ingress controller with single LB, or ClusterIP + Ingress |
| Over-sized PVCs | Paying for unused storage | Right-size PVCs, use volume expansion if supported |
| High replica count in dev/staging | Paying for prod-scale in non-prod | Use Kustomize overlays to reduce replicas per environment |
| CronJobs with large resource requests | Resources allocated for burst but idle most of the time | Right-size to actual job needs, consider spot/preemptible |
Only apply when EKS is detected:
| Check | Impact | Recommendation |
|---|---|---|
| Fargate vs managed nodes | Fargate premium ~20% over EC2 | Use Fargate for bursty/small workloads, managed nodes for steady-state |
| No Spot instances for fault-tolerant workloads | Paying full on-demand price | Add karpenter.sh/capacity-type: spot or node group Spot config |
| Karpenter consolidation not enabled | Nodes not bin-packed after scale-down | Enable consolidation policy in Karpenter provisioner |
| GP2 volumes instead of GP3 | GP2 is 20% more expensive with worse baseline | Use GP3 storage class |
| No cluster autoscaler or Karpenter | Nodes not scaling down when empty | Configure node scaling for non-prod environments |
If you find issues outside the cost domain during the review, recommend the relevant K8s skill:
kubernetes-security-auditkubernetes-operational-reviewkubernetes-helm-reviewsizing-reference.md — Resource sizing guidelines by workload type, common patternsreport-template.md — Full output format template for cost review reportsUse when a skill needs step tracking, progress visibility, and resume support within a single phase. Called by skills, not directly by users.
Use when breaking down a project phase into stories with acceptance criteria, creating ADRs, or planning infrastructure work. Triggers on /plan, story breakdown, detailed plan, architecture doc.
Use when generating, editing, or merging Low-Level Design (LLD) documents at `docs/lld/<component>.md` or `docs/shield/<feature>/lld-<component>.md`. Triggers on /lld command (Path A) and on /plan TRD-driven authoring (Path B, M2 plan). Owns the backend and infra templates, atomic write, provenance stamp, and §14 Changelog convention.
Use when authoring a new PRD or upgrading a lean PRD to standard. Walks user through 20-section problem-first scaffold (or 10-section lean), pre-populates from prior /research transcript if present, defers Terminologies (§2) until after the rest is drafted (auto-fills from research glossary + scan of body), invokes shield:story-coverage between Sections 6 and 8 and shield:milestone-coverage between Sections 8 and 15, prompts for story Type (new/enhancement/existing), supports custom team templates via .shield.json. Triggers on /prd, write a PRD, author a PRD.
Use when the user invokes /backlog or asks to capture, view, promote, or remove an idea from the project backlog at docs/shield/backlog.json. Triggers on /backlog add|view|remove|promote and on agent-side "capture this as a backlog entry" calls during research/PRD/plan/implement flows.
Use when a plan, architecture doc, or execution plan exists and needs expert review before implementation. Produces a scored analysis with a P0-gated verdict and an enhanced plan. Triggers on /plan-review, review my plan, document review.