Post by Jakub Wachocki

AI Day 2026, Roma, Italy. DotNetCode community nailed it. Loved being here. A room of senior tech leaders pushing the boundaries on how to build tech - quietly admitted: the AI workflow we're all celebrating is making us slower. We need to do better, and there are ways. Lots of learning for the Italian community: Marco Caruso , Andrea Belloni , Andrea Saltarello , Antonio Liccardi , Francesco Gallo , Massimo Crippa , Matteo Baccan , Maurizio Moriconi ,Michele Aponte "Your mega-prompt feels productive. It produces credible, wrong code. You don't notice … until production." Then you're debugging a code no human can read, written by a machine no one can question. The teams pulling ahead aren't the ones with the fastest Cursor or Claude Code setup. They're the ones who write specs sharp enough that the agent can't hallucinate around them. Call it context rot. Call it the Mega-Prompt Trap. Call it whatever you want … it's eating engineering productivity from the inside. Five things I'm taking home: → Specs are contracts. Prompts are wishes. → Context rot is invisible until it's expensive. → If the code lives past a month or touches more than one engineer — you cannot vibe it. Period. → The repo, not the chat, is where context belongs. → Code is becoming a derivative of the spec, not the artifact itself. The skill that matters in three years isn't writing code. It isn't even writing prompts. It's the discipline to specify what you actually want - before the machine guesses for you. In Vietbam, we're in the middle of an Agentic AI coding capability assessment across 100+ Vietnamese partners right now. The pattern is already clear. Results out soon How much of your team's time this month went into fixing AI-generated stuff that looked right on the first read?