2.1 Functional and infrastructure requirements
Three layers, in order
| Layer | Question | Example |
|---|---|---|
| Business requirement | What outcome does the organisation want? | "Cut invoice-processing headcount cost by 40% this fiscal year" |
| Functional requirement | What must the system do? | "Extract 14 named fields from PDF invoices and post them to the ERP; flag any invoice where a field is missing or ambiguous" |
| Infrastructure requirement | Under what constraints must it run? | "12,000 invoices/night, finished by 06:00, EU data residency, no PII in logs, $400/month ceiling" |
Items in this skill hand you a business sentence and ask which requirement is missing, or which platform decision it forces. The trap is answering with a technology before the requirement is stated.
Requirement → platform decision
Every infrastructure requirement maps onto a specific, nameable Claude decision.
| Requirement | Decision it forces |
|---|---|
| Results needed now, in a UI | Synchronous Messages API plus streaming so time-to-first-token is low |
| Results needed by tomorrow, in volume | Message Batches API — 50% cost, within 24 h |
| Hard per-request cost ceiling | Model tier plus effort plus prompt caching; measure with usage |
| Hard latency budget | Smaller model, lower effort, shorter output, streaming; fast mode on Opus 5 where available |
| Large stable context reused across requests | Prompt caching with the stable prefix first |
| Very long inputs (whole books, huge repos) | A 1M-context model, plus chunking or context management |
| Machine-consumable output for a downstream system | Structured outputs plus schema validation before the write |
| Data residency in a region | inference_geo on the request, or a provider/region deployment |
| Procurement is locked to a cloud vendor | Claude through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry — see availability tradeoffs |
| Zero/short data retention required | Check per-model availability — some models are not offered under zero data retention |
| Auditability of every action | A workflow with logged steps, not an autonomous agent; plus structured provenance on outputs |
| Irreversible operations in scope | A deterministic gate or human approval, not a prompt instruction |
| Throughput above your rate limit | Batch, request-level concurrency control with backoff, or a priority/committed-capacity arrangement |
Non-functional requirements the guide expects you to elicit
- Volume and shape: requests per day, tokens per request, peak concurrency, input modality (text, images, PDFs).
- Latency: p50 and p95 targets, and whether time-to-first-token or time-to-last-token is the one users feel.
- Cost: per request, per document, per month — and who owns the budget.
- Accuracy target: expressed as a measurable rate on a labelled set, not "it should be accurate".
- Failure behaviour: what happens when the model is wrong, when the API is down, when the schema does not validate.
- Security and privacy: PII classes present, retention, logging rules, who may see model output.
- Compliance: residency, audit trail, human-review obligations.
- Integration surface: which system consumes the output, in what format, with what idempotency guarantees.
"The model should be accurate" is not a requirement — it cannot be tested, so it cannot be met. Convert it: "On the 200-invoice labelled set, field-level extraction accuracy ≥ 98% with zero silent failures; every uncertain field is flagged rather than guessed." That single rewrite dictates the eval, the schema (nullable fields plus a confidence or flag), and the human-review path.
Solution architecture inputs you do not control
Requirements often arrive already constrained by an existing architecture. Read them for what they forbid:
| Constraint in the architecture | Consequence |
|---|---|
| Only outbound HTTPS from the service tier | An MCP server over stdio must run co-located; remote MCP needs a network path |
| No third-party network egress from production | Self-hosted tool execution; no Anthropic-hosted sandbox |
| Existing identity provider and least-privilege model | Tool calls carry the caller's authorization, verified server-side — never the model's assertion |
| Existing queue/worker topology | Batch fits naturally; a long-lived agent session may not |
| Existing observability stack | Log usage, stop_reason, model ID, and request IDs in the format that stack ingests |
Quiz
- Question 1
A stakeholder writes the requirement as "the extraction should be accurate". Which rewrite makes it implementable and testable?
- Question 2
An interactive chat surface sends the same 80,000-token policy corpus with every request. Finance wants the per-request cost down, and the product team will not accept any change to what the model sees. Which mechanism applies?