promptdojo_

The chat is behind glass. The CLI isn't.

Picture how you use Claude in a browser tab today. You ask a question. It writes you something good. Then you do the actual work: you select the answer, copy it, switch to another app, paste it, fix the spots that didn't quite fit. The model did the thinking. You did the moving.

That copy-paste step is the glass. The chat can produce anything but it can't place anything. It has no hands.

The Claude CLI is the same Claude model, running in your terminal, with its hands on the folder you started it in. When you tell the CLI "create a file called notes.txt with these three lines," it doesn't hand you the text to paste. It creates the file. When you say "this folder has five files, rename them to lowercase," it renames them.

Here's the same task, both ways:

ChatCLI
You ask"write me a budget template""create budget.txt with a budget template"
It returnsthe text of a templatethe file budget.txt, sitting in your folder
You thencopy, open an app, paste, savecheck that you like it

Same intelligence. The difference is whether you're left holding the output or whether the output already landed where it belongs.

This matters most when the work is not one file but twenty, not one step but a sequence. Copy-pasting twenty times is where the chat stops being worth it. The CLI doesn't get tired at twenty, or at two hundred.

You don't give anything up by learning the CLI. The chat is still there, still useful, and the next lesson is about keeping both. But the glass is the thing this chapter removes.

read, then continue.