Working with an AI agent

FEWD Studio works with any AI agent that writes to disk. Claude Code is the one we use, and the one most people ask about, but nothing here is specific to it — Cursor, Gemini CLI or anything else that edits files in your repository works the same way.

What a terminal agent is

You open a terminal, go to a project folder, and you are in a session with a model that can see and act on that project.

The important word is act. It is not a chat window handing you a snippet to paste. Within the directory you started it in it can read any file and follow imports across the codebase, create and edit and delete files, run shell commands — your build, your dev server, a git diff, a search — and read the output of those commands and respond to it.

So “the mobile nav is broken below 640px, find it and fix it” is something it can genuinely attempt: grep for the selectors, open the files, make the change, run the build to check nothing broke.

Why this is a different category to page-builder AI

Most people meet AI first inside a tool — a prompt box in WordPress, a generate button in a page builder, an assistant panel in a design app. This is not a bigger version of that. It is a different shape, and it is worth understanding why before you install one.

It has the whole project, not the current block

A prompt box inside a builder sees the element you are editing. It has no idea what your other components look like, what your token file contains, or what naming convention you have been using for six months.

An agent in your project root reads all of it, and can match your existing patterns because it can go and look at them. Keep a standards file in the repo and it reads that every session and works to it — which is exactly what Generate AI files is for.

Its tools are primitives, not a feature list

This is the part people miss. Builder AI can do what the vendor built an action for: generate a section, rewrite this copy, suggest a colour. If what you need is not on that list, it cannot be asked for.

An agent’s tools are the shell and the filesystem. The scope is not a menu, it is anything you could do at a command line. Rename a class across forty files. Write a build script. Read a CSV and generate components from it. Nobody had to ship a feature for any of those.

The output is code you own

Files in your repository, in git, in a format any other tool can read. Not proprietary block data in somebody’s database. Stop using the agent tomorrow and you keep everything it made.

A caveat

If you are not working in a codebase and do not want to be, a prompt box inside the tool you are already in is genuinely the better fit. A terminal agent assumes you are comfortable at a command line and that your work lives in a git repo.

How this fits with FEWD Studio

They solve different halves of the same job, which is why they work together rather than competing.

An agent is strong at reasoning across a whole codebase — refactors, wiring, scripts, anything touching many files at once. It is weak at anything visual. It cannot see your layout.

FEWD Studio is the opposite. The live node tree and round-trip visual editing give you direct manipulation of the thing you are actually looking at, written straight back to disk as real code.

Because both write plain files in the same repository, you move between them freely. Restructure with the agent, then open FEWD and adjust the result by eye. No import step, no sync to manage.

And when the agent gets something wrong, clicking the element and reading the rule responsible beats describing the problem in prose and hoping for a better attempt — see Inspecting an element.

Getting Claude Code installed

You will need a terminal, a supported OS, and a paid Claude account — Pro, Max, Team, Enterprise or an Anthropic Console account. The free Claude.ai plan does not include it.

You do not need Node.js; the installer fetches a self-contained binary.

Anthropic’s own install guide carries the current command for each platform, and is the right place to take it from — install commands and PATH advice change, and a copy of them here would quietly go stale. Claude Code also runs as a desktop app, a VS Code extension and in the browser; the terminal is simply the surface that sits most naturally beside FEWD.

Once it is installed, cd to a project and type claude. FEWD’s own terminal is a perfectly good place to do that, so the agent and the visual editor are looking at the same folder.

Try it on something small first

Ask it to explain a file you already understand. You learn what it is good at much faster than by handing it something important.

On cost

A terminal agent bills against your plan by usage rather than per seat. Large refactors across many files cost more than small edits. Start small, watch what things cost, and scale up once you have a feel for it.