LLMs are here and they’ve already disrupted software development at every level. Revolutions are messy, fraught with challenges, risks and disruption as the status quo evolves to make room for new ideas. Large language models can dramatically improve productivity, but they also confidently output plausible garbage. This session is about getting the speed without surrendering responsibility.
I will be live demoing Claude Code in a terminal by interrogating a real WordPress plugin codebase and working backwards into the workflow that helps make it trustworthy: having Claude itself maintain a CLAUDE.md “project contract”, requiring a plan for major feature development and using a verification loop where changes aren’t pushed unless validation checks pass.
We’ll dig into common points of failure that matter in real projects—hallucinated assumptions, approval fatigue, and the licensing/attribution questions raised when AI tools ingest open source—and how to work in a way that treats the model as untrusted by default.
