How One Dev Team Built a Safer, Smarter AI-Assisted Coding Workflow

A development team has shared the internal AI setup behind their daily engineering work, centered on running AI assistants directly inside a code editor rather than a browser chat window, giving the assistant full access to files, commands, and test results. They run models from two different AI families simultaneously — one to build and one to critically review — exploiting the fact that different model families carry different blind spots. Recurring tasks are codified into structured recipes that embed hard-won institutional knowledge, including past mistakes, so that critical steps are never forgotten or skipped. The team also enforces hard technical blocks — not just text instructions — to prevent high-risk actions like server overwrites or bulk deletions, arguing that a rule without enforcement is merely a note. Their core takeaways are that AI tools should live where the work happens, review must come from a different source than the builder, and past failures must be systematically recorded to prevent repetition.