AI coding assistants are everywhere, but most Drupal developers are either avoiding them entirely or letting them run wild in their codebase. Neither extreme is the answer. This session shows how to use Claude Code to safely build features, add content types and fields, create and update Views, and even fix bugs in contrib modules — all with guardrails that protect your config, your content, and your sanity.
The workflow demonstrated is the same one I use on a ~5,000-page Drupal multisite for the New York State Unified Court System, where accessibility (WCAG AA), security, and config stability are non-negotiable. Because that codebase is proprietary, the live demo will be driven through a purpose-built companion site — **Drupal AI Demo** — that mirrors the same guardrails and patterns in a site we can all poke at freely. You'll see how project-specific **skills** (reusable, versioned instruction sets) encode standards for Twig/preprocess separation, patching contrib modules instead of editing them, config capture workflows, and PHPStan compliance — so the AI follows *our* rules, not its training data.
Key topics:
- Why context is everything**: the "smart zone" vs. "dumb zone," why LLMs are stateless, and why the knowledge cutoff means you can't trust the model's built-in Drupal knowledge
- Claude Code fundamentals**: slash commands, IDE integration, planning mode, `/compact`, memory, and the keyboard shortcuts that actually matter
- Steering the agent**: `CLAUDE.md`, skills, and project rules that make the AI respect your architecture
- Planning before building**: PRD → multi-phase plan → tracer-bullet implementation
- Drupal-specific guardrails**: never editing untracked files, patch workflows for contrib, config capture branches, Single Directory Components, and Views development patterns
- We'll close with a demo of a non-Drupal React Router v7 / TypeScript / Drizzle app where Ralph ships features end-to-end from a multi-phase plan, and talk honestly about where human-in-the-loop is still essential and where you can safely step away.
You'll leave with a concrete mental model for organizing a Drupal codebase for AI collaboration, a starter set of skills you can adapt, and a clear picture of how far this can go today.