When to Use What
Last reviewed
BeginnerWhat you'll learn
~10 min- Use the decision tree to choose the right tool for any task
- Apply the 10-scenario framework to real-world situations
- Identify and correct common misconceptions about tool choice
- Explain why context makes workspace agents better for building, in a desktop app or a terminal
By the end of this lesson, you will be able to look at most common tasks and confidently choose between chat, an in-browser builder, a desktop agent app, and a terminal CLI.
Why this matters
You’ve now run a workspace agent end to end in Lesson 3.1. Here is when chat, a browser builder, a desktop app, or a CLI is the better surface. Plain chat cannot open the folders on your computer or run code in your project; builders and workspace agents can, in different ways.
Knowing which tool to use for which job is the core orchestration skill of this entire course. Get this right, and every module after this one becomes easier.
Four choices, one question
There are four kinds of AI tool to choose from, and they line up from “talks about it” to “works in your project”:
| Choice | Examples | Works on | Best for |
|---|---|---|---|
| Chat mode | ChatGPT, Claude, or Gemini in a browser | The conversation and anything you attach | Questions, brainstorming, learning, writing |
| In-browser builder | Claude Artifacts, ChatGPT Work & Sites, Lovable, Bolt, v0, Replit Agent | Your request and uploaded files, in the provider’s cloud | A quick page, dashboard, or small app to share by link |
| Desktop agent app / IDE agent mode | Claude Desktop (Code tab), ChatGPT Desktop (Codex), Antigravity App, GitHub Copilot App; agent mode in an editor such as VS Code | A project folder on your computer | Building and changing real projects, reviewing each change on screen |
| Terminal CLI / cloud devbox | claude, codex, agy, copilot in a terminal; the same CLIs in GitHub Codespaces or Cloud Shell | A project folder on your computer, or on a cloud machine | Building and changing real projects with keyboard speed, SSH, automation, several agents at once |
The last two are the same agent behind two different doors. A desktop agent app and a terminal CLI from the same maker read the same project folder, follow the same project notes, and have the same plan mode. Choose between them by how you like to work, not by what they can do, and switch whenever you want. (Module 1 covers why many developers end up in the terminal for speed.)
So the key question is not “browser or terminal?” It is “am I thinking, making something quick to share, or building in a project?”
The decision guide
Not sure which to use? Walk through this interactive flowchart:
What are you trying to do?
If the flowchart does not load, here is the quick rule of thumb:
- Does the task involve explanation, brainstorming, or writing? Use chat mode.
- Do you want a quick page, dashboard, or small app to share by link, with no files on your computer? Use an in-browser builder.
- Does the task involve files or a project on your computer? Use a workspace agent: a desktop agent app if you prefer to click and review on screen, or a terminal CLI if you prefer to type (or want SSH, automation, or several agents at once).
- Can’t install anything, but the project is on GitHub? Use a cloud devbox or a cloud agent (GitHub Codespaces, Claude Code on the web). See Lesson 3.4.
Real tasks are often a mix. For example, you might ask chat to explain an error message you don’t understand, then have a workspace agent find and fix it in your project. Tool choice is not always black and white — use your judgment and experiment.
The framework: 10 real scenarios
Here is how the decision plays out in practice. Study this table — it covers the most common situations you will encounter:
| Scenario | Best fit | Also works |
|---|---|---|
| Write an email to a colleague | Chat mode | — |
| Build a webpage from scratch | Workspace agent (desktop app or CLI) | In-browser builder, for a quick one-page draft |
| Learn what an API is | Chat mode | — |
| Debug code that is not working | Workspace agent | Chat, to understand the error message |
| Make a sign-up page for a club event to share by link | In-browser builder | Workspace agent, if you want the files |
| Set up a new project with folders and configs | Workspace agent | Codespaces, on a locked-down machine |
| Compare two technologies | Chat mode | — |
| Edit 5 files to update a feature | Workspace agent | Claude Code on the web, if the project is on GitHub |
| Turn a spreadsheet into a shareable dashboard | In-browser builder | Workspace agent, if the data must stay on your computer |
| Add tests to existing code | Workspace agent | Codespaces or a cloud agent, if the project is on GitHub |
Notice the pattern? If the task involves your files or your project, reach for a workspace agent — in a desktop app or a terminal, whichever you prefer. If you want something quick to share and the files don’t need to live on your computer, an in-browser builder is fastest. If it involves ideas, explanations, or writing, reach for chat.
Why workspace agents are better for building
When you paste code into a chat, there is a lot of friction:
- You copy code from the chat
- You paste it into a file
- You realize you need to copy it into the right file
- Something does not work
- You copy the error message back into the chat
- You get new code
- You copy-paste again
- Repeat 10-50 times
With a workspace agent — desktop app or CLI:
- You describe what you want
- The agent reads your project, makes the changes, and runs the code
- If something breaks, it sees the error and tries to fix it
- You review the result and confirm — this often takes a few iterations
A workspace agent can access the project scope you grant it, which is usually far richer context than pasted snippets. It knows your file structure, your existing code, and your dependencies. This makes it dramatically more effective than copying code back and forth in a browser window.
The biggest advantage of a workspace agent is context. It is like the difference between explaining your house to a contractor over the phone vs. having them walk through it in person. When they can see the actual structure, they build things that fit.
Here is a concrete example. Instead of pasting code into chat and copying the result back, you open your project folder in a desktop agent app — or start the CLI in a terminal in that folder — and describe what you want:
Add a contact form to my website with name, email, and message fieldsThe agent reads your project, creates or edits the right files, and you see the result immediately. (You set one up in Lesson 1.2; Module 6 goes deeper.)
Common misconceptions
“I should use an agent for everything now” No. Chat is still the best tool for conversations, brainstorming, and quick questions. Use the right tool for the job.
“Chat can do the same thing if I paste my code” Technically, yes. Practically, no. The copy-paste workflow is slow and error-prone. Workspace agents see your whole project, run commands, and iterate automatically. It is a completely different experience.
“Desktop agent apps are the beginner version; the terminal is the real thing” No. The desktop app and the CLI from the same maker run the same agent on the same project files. Many developers prefer the terminal for speed, remote servers, and running several agents at once, but a desktop app is a full workspace agent, not a cut-down one.
“Agents are for developers only” The desktop apps need no terminal at all, and the terminal CLIs need only a few commands, which Module 4 teaches. Directing an agent well is exactly the skill you will practise throughout this course.
“I need to understand the code the agent produces” Not line by line, no. You need to understand what it did (created a form, added a route, updated a style) and whether it works (does the page load? does the form submit?). That said, you should review changed files at a high level, run tests when available, and verify behavior before shipping anything. Trust the tool to write code, but trust yourself to check the result.
Practice: pick the right tool
For each scenario below, decide which kind of tool fits best.
You want to understand what 'responsive design' means before building a website. Which tool?
Your website has a bug -- a form submits but nothing happens. You have the project on your computer. Which tool?
You need a one-page sign-up sheet for a club event that people can open from a link. You don't need the files on your computer. Which tool?
You want to add three new pages to an existing website project on your computer and make sure they match the current style. Which tool?
When you are unsure which tool to use, paste this into chat:
Help me decide the right tool for this task: [describe your task]. Should I use plain chat, an in-browser builder, or a workspace agent (desktop app or terminal CLI)? Explain why in 2 lines.
Key takeaways
- Four choices: chat mode, in-browser builder, desktop agent app (or IDE agent mode), and terminal CLI (or a cloud devbox).
- Use chat for questions, brainstorming, learning, and writing tasks.
- Use an in-browser builder for a quick page, dashboard, or small app to share by link.
- Use a workspace agent for building, editing files, debugging, and multi-step project work — in a desktop app or a terminal, whichever you prefer. They are the same agent.
- Context is the workspace agent’s superpower. It reads your project, so it builds things that actually fit.
- There is overlap. Many tasks can be done more than one way. The question is which is faster and more reliable for that specific job.
🔍The evolving AI tool landscapeclick to expand
The line between chat and agents has blurred. Each maker now offers the same agent as a desktop app, a terminal CLI, and a cloud version that works on a GitHub project, and chat apps can build and host small apps in the browser.
The distinction between “thinking tools” and “building tools” still matters, while where the agent runs (desktop app, terminal, or cloud) is increasingly a matter of preference and constraints, not capability.
The skill that does NOT change is knowing what you want to accomplish and describing it clearly. That orchestration ability transfers to any tool, current or future.
Next step: Lesson 3.3 shows how to use plain chat as a second opinion beside your workspace agent.