A practical guide for engineering teams building LLM-powered applications, covering 14 key practices: treating LLMs as systems (not just prompts), defining behavior specs before tuning, versioning prompts and model settings, building evaluation datasets, using appropriate scoring methods, testing tool calls explicitly, setting cost and latency guardrails, monitoring LLM-specific behavior, designing small reviewable prompt changes, using release gates, gradual rollouts, building feedback loops from production failures, assigning clear ownership, and keeping workflows manageable. Includes concrete examples like eval tables with pass/fail gates, judge rubrics, cost budgets per workflow type, and a release checklist.
