Requirements and design with AI

• ADR (Architecture Decision Record) — a short document capturing a design decision, its context, alternatives and consequences.
• Acceptance criteria — testable conditions that define when a story is done.
• Spec-driven development — writing a precise specification first and having the AI implement against it.
• Non-functional requirements — qualities such as latency, security, availability and compliance, as opposed to features.

An LLM is good at structuring: extract stakeholders, goals, assumptions, open questions and acceptance criteria from meeting notes or a one-line ask. The valuable output is the list of open questions, not the polished prose. Have the model separate stated facts from assumed ones so assumptions get confirmed by a person.

Prompt the model to act as a reviewer: list ambiguous terms, missing error and edge cases, conflicting requirements, and unstated non-functional needs. It is cheap to run repeatedly and catches problems before any code exists, where they cost least.

Ask for two or three architecture options with trade-offs against your stated constraints (team size, existing stack, scale, budget). Without constraints, answers are generic. Use the model to argue against your preferred option, and verify claims about specific products against current documentation.

Models draft ADRs, sequence diagrams and Mermaid/PlantUML source quickly from a discussion transcript. The human supplies the decision and rationale; the model supplies the structure. Store ADRs in the repo so future agents and humans can read them as context.

Write a spec (behavior, interfaces, constraints, acceptance criteria), have the agent propose a plan, review the plan, then implement in small steps. Reviewing a plan is far cheaper than reviewing a finished wrong implementation. The spec doubles as the source for tests.

Product priorities, risk appetite, organisational constraints and accountability stay human. The model has no stake in the outcome and can make any option sound reasonable, so a decision made only because “the AI recommended it” is a smell.