https://stijntersteege.com/Short, practical pieces on building software that lasts, and on using AI without the hype.en https://stijntersteege.com/evals-are-the-unit-tests-of-ai-features/https://stijntersteege.com/evals-are-the-unit-tests-of-ai-features/Wed, 21 Oct 2026 08:00:00 GMTHow to test LLM features with a small, honest evaluation set: what to put in it, how to score it, and why it pays off the first time you change a prompt or a model. https://stijntersteege.com/design-your-ai-feature-for-the-wrong-answer/https://stijntersteege.com/design-your-ai-feature-for-the-wrong-answer/Wed, 21 Oct 2026 08:00:00 GMTA better prompt won’t make a language model right all the time. Good AI features are designed around the wrong answer: confidence, review steps, fallbacks and a way back to a human. https://stijntersteege.com/start-with-a-modular-monolith/https://stijntersteege.com/start-with-a-modular-monolith/Wed, 21 Oct 2026 08:00:00 GMTWhy a modular monolith is the right starting architecture for most products: clear module boundaries, one deployment, and an honest path to services when you actually need them. https://stijntersteege.com/review-the-intent-not-the-diff/https://stijntersteege.com/review-the-intent-not-the-diff/Wed, 21 Oct 2026 08:00:00 GMTAI coding assistants make pull requests bigger and smoother. How to change code review so it still catches what matters: intent, boundaries, tests and risk. https://stijntersteege.com/rag-is-a-search-problem-first/https://stijntersteege.com/rag-is-a-search-problem-first/Wed, 21 Oct 2026 08:00:00 GMTMost disappointing retrieval-augmented generation (RAG) projects fail at retrieval, not generation. Practical steps to fix the search part first: clean sources, hybrid search, metadata and measuring recall.