Service capacity and cost planner

What it does

Know how much your service can carry, what each unit of traffic costs, and when you need to scale. You give it your traffic, current resources, measured limits and cost records. It maps demand to capacity, works out headroom, and names the point where you must add more. It also works out cost per unit of demand, sets budget and quota rules, and lists the load tests still needed before you trust any capacity claim. The skill is long, so tell your AI which job you need first. It then uses the matching template.

What goes in and out

You give it

• Traffic/growth/peak estimates, current resources and measured performance.
• Aligned cost records, budget, reliability limits and accountable owners.

You get

• A demand-to-capacity projection with triggers and unit-cost calculations.
• A load/soak evidence plan and an owned budget-versus-reliability decision.

How the task works

• Model demand and headroom. Map request or storage demand to resource limits using justified utilization and scaling assumptions.
• Calculate comparable costs. Align periods, scope and allocation methods; inspect tenant distributions when averages hide skew.
• Gate the capacity claim. Define representative load/soak acceptance evidence and escalate unresolved budget or reliability tradeoffs.

Full instructions

• Start here. Paste only the job you need, such as a growth projection, a peak-event plan, a unit cost, a budget cap or a load test plan. Reply with your inputs when your AI asks. You do not need to paste the whole skill at once.
• Gather your traffic history, current resource counts, and a cost export for the same period.
• Add your reliability targets, budget cap and the names of who owns each decision.
• Open Claude, ChatGPT or Gemini and paste in the skill. No code or paid tool is needed.
• Paste your inputs and ask for a capacity and cost plan. Give your AI the matching template from the skill folder: capacity-model.md for a growth projection, load-soak-test-plan.md for tests, unit-economics-record.md for unit cost, or budget-quota-decision.md for spend rules.
• Check the flagged gaps. Any claim marked as untested needs a real load test before you rely on it.
• Next, find waste in the bill with the cloud bill waste finder, and set your uptime target with the reliability target setter.

Words used in this task

• headroom: How much extra traffic your service can take before it hits its limit.
• load test and soak test: A load test sends heavy traffic to find the limit. A soak test keeps steady traffic on for hours or days to find slow leaks.
• error budget: The amount of failure your uptime target allows before you must stop and fix things.
• hot tenant skew: One customer using far more than the rest, so the average hides the real load.
• pooled or siloed: Pooled means customers share the same resources. Siloed means each gets their own.

Full instruction file

Capacity and Cost Engineering

Connect demand, performance, reliability, and spend into defensible capacity and cost decisions. This skill models technical capacity, calculates unit cost at the infrastructure level, defines budget and quota controls, requires load/soak evidence for capacity claims, and makes cost-performance tradeoffs explicit — producing evidence that feeds production-readiness (source reference: ../production-readiness/SKILL.md) launch decisions and constrains or supports site-reliability-engineering (source reference: ../site-reliability-engineering/SKILL.md) SLO choices.

Before you start

Use the material below to do this job. If facts or tool access are missing, explain what you can finish and what must wait.

What you need (use your own material or the linked skill):

• Your question (yours): the decision or problem, and what a good outcome looks like.
• Your data (yours): the data files, with their format and where they came from.

How to start:

1. Look first. Use what they've already attached, pasted or put in their about-me note. In a coding assistant with file access, look in the project too for these by name.
2. Check what's missing. Look for your question, your data. They can attach files, paste the details, or share a file path your assistant can read. Ask only for missing facts that change the result. If they ask you to start with what they have, complete the parts you can and mark the gaps.
3. Do the job. Follow the skill below with the facts and tools you have. Mark each gap with MISSING: where it belongs and never invent a fact to fill it.

Connected dimensions

Demand (traffic and growth), performance (latency and throughput), reliability (SLOs and error budgets), and spend (cost) are treated as connected dimensions — a change in any one dimension affects the others. The skill's core method is tracing the connection:

A capacity decision that changes one dimension without modeling the others is incomplete. Every capacity model, unit-cost calculation, and budget decision in this skill must name at least one connection to another dimension with evidence or an explicit assumption.

Loading Guide

Load this skill when the task involves any of:

Working method

1. Model demand and capacity

Start with the demand signal: current traffic, growth rate, and any known peak events. Translate demand into capacity requirements using a capacity model that connects:

• Demand units (requests/second, concurrent users, GB ingested, messages/second) to
• Capacity units (vCPUs, memory GB, storage GB, IOPS, network throughput, provisioned throughput units) through
• Utilization targets (maximum sustainable utilization per resource, stated with rationale — never a universal percentage without context).

A capacity model is incomplete without a stated utilization target, the evidence for that target (why 70% and not 85%?), and the scaling trigger that fires when utilization approaches the target.

Use the capacity model template (templates/capacity-model.md) which captures demand assumptions, capacity-unit mapping, utilization targets with rationale, scaling triggers, and a projection over the relevant horizon. The template requires fields for assumptions, evidence source, ownership, and tradeoffs.

2. Calculate unit cost

Unit cost is the cost of serving one unit of demand — cost per request, cost per user per month, cost per GB stored, cost per provisioned capacity unit. Unit cost connects infrastructure spend to product and reliability decisions.

Calculate unit cost as:

unit cost = total cost of capacity / number of demand units served

Both numerator and denominator must be measured over the same period, with the same scope (service, team, or platform), and with the same allocation method stated (direct resource cost, attributed shared cost, or fully loaded cost including overhead).

The unit-economics record template (templates/unit-economics-record.md) requires: the unit definition, the cost numerator with allocation method, the demand denominator with measurement source, the resulting unit cost, a cost-per-SLO comparison (what happens to unit cost at 99.9% vs 99.99%?), and an assumptions/evidence/ownership/tradeoffs section. The template includes a structured field for the unit-cost calculation formula.

For a multi-tenant service, do not use a fleet average as the only unit. Load references/multi-tenant-capacity-and-unit-cost.md and distinguish platform baseline cost from tenant-variable cost, then report a distribution of tenant costs or resource consumption. A tenant's variable unit cost may depend on request mix, storage, background work, burst shape, placement, and tier entitlements. Shared-cost allocation is an explicit modeling choice, not a claim that every tenant consumes an equal share.

3. Define budget and quota controls

Budget controls are spending limits with operational consequences — spending alerts at thresholds, hard caps that prevent further spend, and the operational behavior when a cap is hit (degrade, throttle, or stop). Quota and rate-limit enforcement are the mechanisms that implement budget controls at the request or resource level.

Budget controls operate at three levels:

Budget controls must specify what happens at each threshold — the operational behavior, the user-facing impact, and the owner accountable for responding. A budget threshold without a defined operational consequence is incomplete.

The budget/quota decision template (templates/budget-quota-decision.md) captures: budget owner, period, thresholds (alert/soft/hard), quota or rate-limit configuration, enforcement mechanism, operational behavior at each threshold, cost attribution method, anomaly detection triggers, and approval record.

4. Require load/soak evidence

Load and soak test evidence is required for capacity decisions. A capacity model alone — without observed system behavior under representative load — is insufficient evidence for a capacity claim or a scaling policy.

• A load test exercises the system at a target throughput (e.g., expected peak + 20% headroom) for a defined duration and measures latency, error rate, and resource utilization.
• A soak test exercises the system at a sustained load over an extended period (hours to days) and detects slow leaks (memory, file descriptors, connection pools, disk growth) that a short load test misses.

The load/soak test plan template (templates/load-soak-test-plan.md) captures: test objective, target throughput with rationale, duration, environment (must be representative — a dev-environment test is not sufficient), success criteria (latency percentiles, error rate, resource utilization), data collection plan, and the evidence record. The template distinguishes a component-level benchmark from an end-to-end test; a capacity decision must state which boundary was exercised.

Modeling without test evidence, or testing without a model, is incomplete. Both are required.

For multi-tenant claims, the evidence must exercise representative tenant profiles together, including ordinary tenants, high-demand tenants, bursty tenants, and relevant tier or placement variants. A single-tenant benchmark or fleet-average test cannot establish protection against hot tenants, partition skew, or fairness behavior.

5. Resolve SLO-cost tradeoffs

An SLO-cost tradeoff arises when the cost of meeting an SLO at projected demand exceeds the budget, or when a budget constraint forces a lower SLO than the team would otherwise target. This is a structured decision, not an implicit acceptance.

The SLO-cost tradeoff record template (templates/slo-cost-tradeoff-record.md) captures: the SLO under discussion, current cost to meet it, projected cost at demand forecast, alternative SLO with cost comparison, degradation path if the lower SLO is chosen, error budget impact, accountable owner, and approval record. The template requires surfacing the tradeoff with evidence (cost projection, load-test data) and ownership (who decides and who is accountable).

Scenarios

Growth scenario

Demand is growing predictably (e.g., 15% month-over-month). The question: when does current capacity become insufficient, and what does it cost to stay ahead of growth?

Guidance:

1. Project demand forward using the growth rate, with confidence intervals.
2. Model capacity at current utilization targets; identify the resource that saturates first.
3. Calculate the cost of incremental capacity at each scaling step.
4. Define the scaling trigger — the utilization threshold at which provisioning must begin (lead time matters).
5. Record assumptions (growth rate stability, no step-change events, current utilization pattern holds) and evidence sources.

Peak scenario

A known event will drive traffic well above baseline (product launch, Black Friday, seasonal peak). The question: how much capacity is needed for the peak, what does it cost, and is the cost justified?

Guidance:

1. Model peak demand separately from baseline — peak shape (height, duration, ramp), not just the peak number.
2. Size capacity for the peak, not the average. Include headroom.
3. Run a load test at the projected peak throughput before the event. A model without load evidence is insufficient.
4. Define the post-peak scale-down plan and its trigger — capacity that persists after the peak incurs unnecessary cost.
5. If peak capacity cost exceeds budget, model a degraded peak alternative (which functions shed, what users experience).

Degraded scenario

A dependency fails or a resource constraint forces operation below full capacity. The question: what does the system look like in degraded mode, what capacity is needed for the core path, and what does degraded operation cost?

Guidance:

1. Tier functions: core (must preserve) vs enhancing (can shed). Route degradation-path design to resilience-and-recovery (source reference: ../resilience-and-recovery/SKILL.md).
2. Model capacity for core-path-only operation — what resources are freed by shedding enhancing functions.
3. Calculate the cost of degraded operation (may be lower than full operation, or higher if failover resources activate).
4. Define the maximum degraded-operation window — how long degraded mode can persist before escalation.
5. Ensure the degraded capacity model is exercised in a game day or chaos test (route to resilience-and-recovery for exercise design).

Cost-constrained scenario

A budget constraint prevents provisioning to the ideal capacity or SLO target. The question: what is the best achievable reliability and performance within the budget, and who decides?

Guidance:

1. Start from the budget constraint — state the cap explicitly (monthly, quarterly, annual).
2. Model the capacity that the budget can purchase at current unit costs.
3. Calculate the SLO and performance targets that capacity can support.
4. Compare to the unconstrained ideal: what SLO, latency, and throughput are being traded away.
5. Produce an SLO-cost tradeoff record with an accountable owner. Cost optimization must not degrade reliability, privacy, or user outcomes — if the budget cannot support an acceptable SLO, the decision is escalated, not silently accepted.

Cost optimization must not justify degrading reliability, privacy, or user outcomes. If a cost constraint forces a choice between budget and these non-negotiables, the tradeoff is escalated to an accountable owner with the evidence — it is never silently accepted as an optimization.

When not to use

This skill does not provide generic cloud-cost tips (reserved instances, spot instances, "turn off unused resources," "right-size," savings plans). Those are platform-specific implementation tactics that belong in platform-engineering (source reference: ../platform-engineering/SKILL.md) references or cloud-provider documentation, not in a capacity-and-cost methodology skill. This skill owns the decision framework and evidence standard for capacity and cost — not a list of cost-cutting tips.

This skill does not prescribe universal utilization targets. A utilization target of 70% for a latency-sensitive service with spiky traffic is not the same as 85% for a batch-processing pipeline with predictable load. Every utilization target must be stated with context, rationale, and the evidence that supports it. "Target 70% utilization" without context is not a capacity decision — it is a guess.

This skill does not permit cost optimization to justify degrading reliability, privacy, or user outcomes. These are non-negotiable constraints. A cost-constrained scenario that would violate them must be escalated, not optimized around.

Routing table

File map

After this

When you've handed it back, suggest what they could do next in one short question. It's a suggestion: if they're done, that's fine. Offer these, and say they can type their own:

• The AWS cost waste finder.
• The Service reliability target planner. If they pick one and that skill is added here, carry straight on with what you just made. If not, give the link and say what to bring.