Engineer AI systems, not model calls.

Define the discipline, explain why AI systems differ from deterministic software, and establish the language used throughout the handbook.

Translate user needs into AI-appropriate requirements, decide whether AI is justified, and identify the blast radius before implementation.

Introduce the architecture model that connects business problem, engineering problem, evaluation, harness, context, prompt, model, and infrastructure.

Move beyond prompt tricks into durable instruction design, prompt contracts, examples, structured outputs, and prompt change management.

Design the data, retrieval, ranking, memory, compression, and authorization layer that makes AI behavior grounded and useful.

Build the operating layer around the model: routing, retries, validation, fallbacks, state transitions, approvals, and traceability.

Address prompt injection, data leakage, model and supply-chain risk, privacy boundaries, compliance, retention, and responsible AI review.

Safely connect AI systems to tools, APIs, files, databases, workflows, and long-running tasks without losing control of authority.

Package, deploy, version, route, scale, and roll back AI systems across model versions, prompts, retrieval indexes, tools, and policies.

Compare common patterns including RAG, copilots, workflow agents, AI search, human-in-the-loop systems, and hybrid architectures.

Create the quality system for AI: scenario suites, golden datasets, adversarial tests, human review, production sampling, and release gates.

Operate AI systems using traces, logs, semantic quality signals, cost telemetry, drift monitoring, incident response, and continuous improvement.

Walk through realistic systems such as customer support, enterprise search, document intelligence, financial copilots, coding assistants, and operations agents.

Provide reusable operating assets that teams can copy into real projects and adapt for their own review process.