agentleFS
imMamdouhaboammarimMamdouhaboammar/anti-slop-engineering/skills/finops-and-unit-economics-governor/SKILL.md

finops-and-unit-economics-governor

Skill4 starsChanged 31 days agoMITView on GitHub

Review cost per request, job, or tenant across cloud egress, API fan-out, retries, GPU usage, and model tokens. Use when an input or dependency failure can multiply paid work or when a proposed change cites cost savings.

The file

Files
SKILL.md

FinOps and Unit Economics Governor

Review a billable unit

  1. Select the outcome to count, such as a completed export. Count accepted, failed, and cancelled work separately. Record the observation window, region, provider, unit of measure, and workload segment.
  2. Trace the request through billable fan-out, dependency retries, queue redelivery, egress, GPU duration, and model input/output/cache units. Include failed and speculative attempts when billable. Reconcile the trace with billing export or provider usage records.
  3. Build base and bounded stress worksheets. For each meter record quantity per accepted request, worst permitted quantity, rate, currency, billing unit, source, and effective date. Keep provider-specific rates outside the shared example. Convert bytes, milliseconds, and token blocks before multiplying, respecting rounding and minimums. Missing rates remain unknown.
  4. Bound work at admission and before each paid downstream call. Verify per-tenant and shared limits, total attempt budget including the first call, queues or semaphore, request deadline, and cancellation. Determine who can raise caps. Replays should reuse the underlying result where the paid side effect can be deduplicated. A concurrency limit controls how rapidly work starts. A bound on units or attempts controls how much work one request can generate.
  5. Compare budgets against measured spend per outcome and per tenant. Record window or rate limits, billing lag, and behavior when the meter is unavailable. Before changing paid controls, test a safe rejection or lower-cost fallback and confirm the user-visible result.
EvidenceWhat to measure
API fan-outCalls per request by dependency and attempt number, context propagation, caller deadlines, and tail distribution
NetworkBytes by source/destination and region or zone, cache hit/miss, retries, and provider billing unit
Model or GPUUncached input, cached input, output units, billed cancellations, GPU seconds, concurrent slots
Retry and failureNumber of paid attempts before a terminal outcome, queue age, deduplicated effects, and user-visible outcomes

Worksheet and fixtures

For each meter, multiply billable units by the applicable rate. When relevant, add fixed charges with an explicit allocation period and denominator. Sum only after normalizing units and currency. Label an estimate from a run with missing prices as incomplete, even when a partial total is known.

The offline cost cases use synthetic USD rates. They distinguish a bounded fan-out and retry budget, observed work without an upper bound, and an unavailable price. The fixture consumes billable counts already obtained from a trace. It does not query a provider or forecast spend.

Catalog: CON-09, OPS-07, SRE-08.

References: FinOps unit economics, FOCUS specification, Google SRE on managing overload.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.