Recipes for building with generative AI.
Practical, opinionated patterns that hold up in production — agent harnesses, spec-driven development, memory, and prompts. Each one comes with a diagram and runnable code.
Architecting an AI-native product
Most 'AI features' are a chatbot bolted onto a CRUD app. An AI-native product is built the other way around — capabilities first, with the model in the critical path and the UI, data, and org reshaped to match. Here's the reference architecture.
Build an effective agent harness
An agent is a loop around a model, and the loop is the easy part. This is a deep dive into the part that actually matters: the guardrails, tracing, and control flow that let you leave it running.
An evaluation harness you can ship on
You can't improve — or safely ship — what you can't measure. This is how enterprises turn 'the demo looked good' into a regression-gated eval suite that tells you, before deploy, whether a change made the product better or worse.
Spec-driven development with AI agents
Code is cheap now; intent is the bottleneck. Write a tight spec, let the agent implement against it, and gate the work with executable acceptance checks — so you review intent instead of archaeology.