Tommaso Nervegna, a Design Director at Accenture Song, gives one of the clearest practitioner accounts I’ve seen of what using Claude Code as a designer looks like day to day.

The guide is detailed—installation steps, terminal commands, deployment. This is essential reading for any designer interested in Claude Code. But for me, the interesting part isn’t the how-to. It’s his argument that raw AI coding tools aren’t enough without structure on top:

Claude Code is powerful, but without proper context engineering, it degrades as the conversation gets longer.

Anyone who’s used these tools seriously has experienced this. You start a session and the output is sharp. Forty minutes in, it’s forgotten your constraints and is hallucinating component names. Nervegna uses a meta-prompting framework called Get Shit Done that breaks work into phases with fresh contexts—research, planning, execution, verification—each getting its own 200K token window. No accumulated garbage.

The framework ends up looking a lot like good design process applied to AI:

Instead of immediately generating code, it asks:

“What happens when there’s no data to display?” “Should this work on mobile?” “What’s the error state look like?” “How do users undo this action?”

Those are the questions a senior designer asks in a review. Nervegna calls it “spec-driven development,” but it’s really the discipline of defining the problem before jumping to solutions—something our profession has always preached and often ignored when deadlines hit.

Nervegna again:

This is spec-driven development, but the spec is generated through conversation, not written in Jira by a project manager.

The specification work that used to live in PRDs and handoff docs is happening conversationally now, between a designer and an AI agent. The designer’s value is in the questions asked before any code gets written.