Lessons from building software with AI — practical, experience-based, no fluff.
AI agents ship features in hours. If your test code isn't bigger than your production code, they're outrunning your safety net. Here's what 100% coverage actually looks like.
615 passing tests, full coverage, and I still couldn't debug a failure without reading the source code. The Screenplay pattern fixed that.
Practice videos were getting lost in a WhatsApp group. I needed to track what I had practiced. So I built a web app — 9 commits, 414 tests, 1 developer, 10 AI agents, 1 day.
My Swedish teacher app needed to decide if each message was a translation attempt or a conversation. Four string prefixes kept breaking. One LLM call fixed it permanently.
Why separating architecture, quality, and security from code generation produces better AI-assisted software — and mirrors how professional teams actually work.
How Domain-Driven Design and Ubiquitous Language became the most effective way to get accurate code from AI — no clever prompts required.
As a Java developer moving into Python through AI, I found the testing stack I trusted did not exist. So I built screenwright — the Screenplay pattern for Python with cinematic BDD reports.
Infrastructure to backend to API contract to frontend — each handoff is where context gets lost and bugs creep in. AI collapses all four into a single session.
AI can generate code and tests in seconds — but how do you know the tests are actually verifying anything? Mutation testing is one powerful way to find out.