agentsclimarketplace
Sign in

Cloud cost optimization

Skill Amey-Thakur/AI-SKILLS/skills/cloud/cloud-cost-optimization

Cut cloud spend with tagging, rightsizing, commitment mix, and egress awareness, without breaking reliability. Use when the cloud bill needs reducing or a cost-review practice needs standing up.From its SKILL.md

Install
npx -y skills add Amey-Thakur/AI-SKILLS --skill cloud-cost-optimization

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.7 KB, 606 tokens by cl100k_base, as published. Nobody here has run it

Cloud cost optimization

You cannot cut what you cannot attribute. Visibility first, then the big three levers: rightsizing, commitments, and storage lifecycle; everything else is decoration until those are done.

Method

  1. Attribute before optimizing. Enforce tags (team, service, env) at provision time via IaC policy, not wiki pleading; untagged resources get a weekly report and an owner hunt. Turn on the billing export to a queryable store; the console's month view hides the per-service trends you need.
  2. Rightsize from utilization, not fear. Pull 30 days of p95 CPU/memory per instance; anything under ~40% at p95 drops a size (halving cost per step). Do the same for provisioned IOPS, over-replicated dev databases, and idle load balancers. Schedule non-prod to sleep nights and weekends: 70% of the week is off-hours.
  3. Buy commitments for the floor, spot for the burst. Cover the stable baseline (12 months of history says what that is) with savings plans/reserved capacity at 60-70% coverage; leave headroom on-demand. Fault-tolerant batch and CI go on spot at 60-90% off, with interruption handling tested (see kubernetes-workloads for disruption budgets).
  4. Lifecycle the storage. Objects: transition to infrequent-access after 30 days, archive after 90, delete what compliance does not require (see data-retention); abort incomplete multipart uploads. Snapshots and orphaned volumes/IPs are the classic silent leak: sweep monthly.
  5. Respect egress and cross-zone gravity. Data leaving the provider or crossing regions costs more than compute reading it in place. Process where data lives, cache at the edge, and check NAT gateway processing charges: a chatty service behind NAT can out-cost its instances.
  6. Institutionalize the loop. Unit economics metric (cost per request/customer/GB) on a dashboard, anomaly alerts on daily spend, a monthly 30-minute review per team with the top-5 movers, and cost as a line item in design reviews (see architecture-review-board). One-off cleanups decay in a quarter; the loop does not.

Boundaries

  • Do not trade reliability invisibly: dropping multi-AZ, backups, or headroom is a risk decision for the service owner, not a cost optimizer's line edit.
  • Engineering time is a cost too; a week of work to save $40/month fails its own review.
  • Provider pricing shifts and this skill's ratios drift; re-verify numbers against current price sheets before big commitments.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most token cost skills give in 606 tokens

Counted across 84 of the 95 authors here whose files we hold, read 2026-09-06

  • Set explicit budget limits before processing batchesin 14 of 84, across 9 files
  • Define model names as constants or configin 14 of 84, across 9 files
  • Route simple tasks to cheaper modelsin 14 of 84, across 9 files
  • Log model selection decisionsin 14 of 84, across 9 files
  • Retry only transient errorsin 14 of 84, across 9 files
  • Track spend with immutable cost recordsin 13 of 84, across 8 files
  • Cache system prompts over 1024 tokensin 12 of 84, across 7 files
  • Check the budget before each API callin 12 of 84, across 7 files
  • Start with the cheapest modelin 9 of 84, across 8 files
  • Back off exponentially between retriesin 9 of 84, across 8 files
  • Review costs weeklyin 8 of 84, across 6 files
  • Use reserved capacity for steady workloadsin 8 of 84, across 6 files

Said here and by no other author read

  • Enforce resource tagging at provision time via IaC policy
  • Export billing data to a queryable store
  • Rightsize instances from 30-day p95 utilization
  • Schedule non-production resources to sleep off-hours
  • Cover the stable baseline with savings plans
  • Lifecycle objects to infrequent-access then archive tiers

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.