AI in product

Most AI features fail in production not because the model was wrong — but because nobody asked the right questions at scoping.

agentmemory is an open-source tool that lets Claude Code, Cursor, Copilot, and other agents remember your project across sessions. Here's how to run it.

What the user sees when the AI is wrong is more important than what they see when it's right. Spec the fallback first.

Curated test sets always perform better than production data. If you're not testing on production-shaped inputs, your benchmark is a press release.

Models drift, providers throttle, prompts break. If your feature has no plan for what to do when the AI is unavailable, you don't have a production feature.

The cases where adding AI actively makes the product worse. Recognize them before you spec.

Every AI feature should be tested against one question: does this remove a step the user is already taking? If no, kill it.

Build evals when wrong outputs cost money, customers, or trust. Skip them when you can fix a bad output in five minutes.

Most teams that 'need' a vector database actually need keyword search and a small JSON file.