Part 4: Claude Code Arrives, and the AI Gets a Job Title

Copying code out of a chat window and pasting it into files had cost me plenty in Part 3. So on September 1 I installed Claude Code, a version of Claude that works directly in my project folder from the terminal. It reads the files, makes the edits and runs the commands itself.

It didn't start smoothly. After the install, typing claude gave "command not found." The program was there; my Mac just didn't know where to look for it. Adding its folder (~/.local/bin) to my PATH, the list of places the terminal searches for programs, fixed it.

Two tools, two roles

We set a clear split. Claude Code edits the files, like a programmer at the terminal. The chat handles planning, design decisions and database work, like the systems analyst who writes the spec. Anyone who has worked on a team with those two roles will see why this helped. Each side knows its job, and nobody steps on anyone else's work.

Three weeks of building

With that split in place, the work moved fast. The six Workshop rooms were renamed and restyled, and new Latest News, Workshop Equipment, Behind the Scenes and Front Desk pages were built. I added a mobile menu, a new navigation order with a Store placeholder, footer social links, an editable Info Strip, Google Analytics, a forgot-password flow, an Events feature with a ticker on the home page, the Privacy Policy page, newsletter signup, and the project blog entries you're reading right now.

I also made some business decisions. Netlify stayed as the host on its $9-a-month plan for now, and GoDaddy stayed as the domain registrar. Since I use GoDaddy email on the domain, I kept GoDaddy's DNS and added two records pointing the site to Netlify. Claude advised against GoDaddy's masked forwarding, which would have broken logins, security certificates and search ranking. Then came my favorite message of the month: "domain complete and tested."

Production-only surprises

The home page's hero text and featured picks were being fetched from Sanity but never actually handed to the page, so edits in Studio did nothing. More worrying, the newsletter signup worked on my computer but failed on the live site. Next.js hides error messages in production for security, so the real error never reached the screen. That set a rule for the rest of the build: functions return a result saying whether they worked, rather than throwing an error. Separately, Sanity's command for publishing schema changes failed until a missing settings file, sanity.cli.ts, was created.

The lesson

Give each AI tool a defined role, the way you'd staff a project with a programmer and an analyst. And test on the live site early, because some failures only appear there.

Try it yourself

If you're juggling more than one AI tool, write down in one sentence what each one is for. When a feature works on your computer, push it live and try it there before you call it done.

Next time: an upgrade that didn't happen, and why that was the right outcome.