Community

finops-review Agent Skill

Open SKILL.md package · coding · Claude Code, Codex, Cursor, OpenCode

Review cloud spend and FinOps plans — idle resources, rightsizing, commitment discounts, and unit-cost metrics. Use whenever the user shares a bill, Cost Explorer export, asks why AWS/GCP/Azure is expensive, wants a savings plan, or reviews infra for cost before scaling.

View on GitHub

FinOps Review

Treat spend as a product metric. Cut waste first, then buy commitments, then redesign architecture. Do not recommend reserved capacity on a workload you have not seen stabilize.

For IaC blast radius, use terraform-review / pulumi-review / cloudformation-review. For load that drives cost, use load-test-planning.

Workflow

  1. Scope: account/project, env, time window, currency, who owns the bill.
  2. Split the bill: compute, storage, data transfer, managed DB, logs, support, marketplace. Name the top 5 line items.
  3. Classify each: idle / oversized / correctly busy / unknown.
  4. Unit cost: spend ÷ a real driver (active users, requests, GB stored, jobs). A cheaper month with 2× traffic is a win.
  5. Quick wins: unattached volumes, idle load balancers, unused IPs, forgotten snapshots, always-on non-prod, log ingest without retention caps, NAT/egress surprises.
  6. Rightsizing: use observed CPU/mem/IOPS, not instance family folklore.
  7. Commitments (RI / SP / CUD) only after 30+ days of stable baseline and an owner who will use them.
  8. Guardrails: budgets, anomaly alerts, tags/labels required to ship.

Severity tags

  • [blocking] prod data-loss path proposed as a saving (delete the only backup, turn off Multi-AZ with no RPO)
  • [important] idle spend >10% of the bill, missing tags, uncapped logs/egress, non-prod = prod size 24/7
  • [nit] naming, dashboard cosmetics

Output format

## FinOps review: <account / window>

**Total:** …
**Unit cost:** … / <driver>
**Top drivers:** …

### Findings
1. [important] …

### Savings backlog
| Item | Est. monthly | Effort | Risk | Owner |

### Do not do
…

### Measurement
budget, anomaly alert, tag policy

Rules

  1. Never invent dollar amounts. If the bill is missing, give relative actions and ask for the export.
  2. Do not recommend deleting backups, dropping Multi-AZ, or turning off encryption to save money.
  3. Separate one-time cleanup from recurring architecture cost.
  4. Tag/label gaps are findings: unallocated spend cannot be owned.
  5. Egress and cross-AZ traffic are first-class; "compute looks fine" is not a complete review.
  6. Spot/preemptible only for interruptible work, with a fallback.
  7. Distinguish list price from the user's actual discount/EDP if shown.

Edge cases

  • "Make it 50% cheaper": refuse a number without a bill; offer the measurement + top-waste hunt instead.
  • K8s / shared clusters: allocate by namespace/label; do not blame the control plane for app replicas.
  • AI / GPU: utilization and batching beat instance shopping.
  • FinOps theater: a pretty dashboard with no owner and no budget page is not a control.

---