FROM CHAT WINDOW TO COMMAND LINE

All right. Hey there. I'm Joe Amditis. I am Associate Director of Operations at the Center for Cooperative Media at Montclair State University, and this is Advanced Prompt Engineering for Journalists.

This is a follow-up to my earlier course from last year, which focused on beginner prompt engineering principles, tactics, and tools. And this course is primarily about getting you out of the web chat interface, where you open the browser and go to chatgpt.com and use your custom GPTs. We're gonna escape that. We're gonna get out of the browser and bring these tools onto your local machine where they can interact with, edit, create, or delete, in some cases, real files and folders.

And I know some of you, uh, are probably using company machines or laptops or computers where you don't even have access to that kind of, uh, admin privileges, and the IT department or IT guy that you frequently interact with would probably lose his mind, uh, if you ask them for a direct route access into the shell of a company machine. So fear not, if you are unable to or unwilling to, uh, use the terminal or command line, um, there are other options. You can use your favorite IDE, your favorite integrated development environment, like VS Code, Antigravity, if you're one of those people, Cursor, Windsurf. There's a million of these things out here.

Um, so whichever one you like best, as long as you're able to interact with a large language model, in particular a paid subscription or API usage version of either Gemini, ChatGPT, Codex, uh, or Claude and Claude Code. As long as you have access to one of those, I don't care if you use it in Antigravity or you're using it in the terminal, just know that I will be using PowerShell primarily. Um, and so any of the things that I do or things that I explain will be geared towards that modality. I do have a bunch of resources on the, uh, MOOC website that we, uh, I created for this course, mooc.amditis.tech.

Uh, I also have tools.amditis.tech. I'll put both of the links on the screen right here, uh, and I'll make sure that that is kept up to date for you to browse and copy or fork or whatever. But that's where, how we're gonna be working with this stuff. Another, uh, platform that we're gonna be using that is not necessarily required but highly recommended if you don't already have one, is a GitHub account.

Personally, I've spent about the last year really burying myself in GitHub and trying to understand how it can be useful, and, uh, it, it's only somewhat recently that a lot of the things have started to click. Uh, I'm gonna be talking and approaching these things from the perspective of someone who is completely, almost completely self-taught on this. And, uh, the reason why I feel qualified to do that is because I think a lot of you are also in a similar position, uh, where you don't have the time, money, or attention span maybe to, uh, to enroll or invest in yourself in that way and get a full software engineering degree, and that's fine. Um, we're gonna be focusing on a lot of the stuff in this course from the perspective of someone who is not an expert software engineer.

So if you are, and you notice something that I say is either wrong or could be bet- phrased better or there's a better solution, go to the instructor forums. Go to the discussion forums. Uh, let me know. We're gonna talk about all that stuff as we go on, but the f- the main thing that you should know about this is you need a paid account for one of the three major models.

That's ChatGPT or Codex, Gemini, Gemini CLI, or Claude and Claude Code from Anthropic. The twenty dollar a month plan is fine. If you decide to use the API, that's also fine, although I've done the calculations on this, and it's very obvious that it's way more expensive to do the cost-by-use API style, uh, subscription, but if that's what you have access to, that's fine. The only thing is you will not be able to do most of the things that we are doing here with a free account.

The only exception is possibly Gemini, uh, which does have a free plan that allows you to use their CLI agent, um, in whatever tool you're deciding to use. There are also other, uh, tools that you can use that are still browser-based but do somewhat similar functions, things like, uh, the Google AI Studio, Replit, Lovable, and if you've used any of those tools already, um, you probably have some idea of what I'm talking about. But for what we're doing, I'm gonna be showing you PowerShell stuff, mostly with Claude Code. But what you need to know for now is we're escaping the browser.

We're getting out of the browser. We're getting onto our machines. And the reason is that these tools are much, much more powerful, and you can accomplish and achieve and build a lot more and more useful things if you are able to take advantage of the processing power and the ability to run scripts and workflows, schedule cron jobs and recurring tasks, uh, scrapers, letting something run overnight without having to be mediated through the web browser. Um, so that's the value here.

That's why we're, we're so focused on getting you out of the browser, even though we will return to the browser for various things, one-off tasks, maybe even to manage remote sessions on the Claude web app from your phone on your mobile browser. But, uh, the majority of what we're doing is gonna be in the, in the PowerShell or in the terminal. When it comes to GitHub, GitHub is gonna serve as our sort of, um, a very simplified, maybe oversimplified comparison, but a sort of like our Google Drive for our code. And I say that specifically because before I knew about GitHub and before I became familiar with it over the last year, uh, I straight up was using Google Drive to store and sync up my code.

And what it resulted in was a nightmare of version history mismatches and version control nightmares and rewrites and overwrites, um, you know, accidental deletions. It was a nightmare, and, uh, GitHub, uh, saved me from all of those nightmares. So if you are unfamiliar with it or if you're really familiar with it, uh, it's important. We're gonna be using it, uh, even though not all of our projects necessarily would require GitHub.

So why don't we pause there, and let's start at the beginning, and let's get you set up with one of your agentic CLI agents on your own machine. All right. Now, before we officially begin, there are a few core concepts that I think it's important that we understand or at least touch on before we jump in, and one of them is why we want to escape the chat window of the browser. Um, and for this, I recommend reading Ethan Mollick's piece on the harnesses that the AI tools like this use.

Um, but essentially, it boils down to a sort of an argument that the tools you're interacting with, the models, the large language models, um, they behave differently based on the environment in which they are operating. And the different harness that you give this model impacts its capabilities and its reactions, but also the, the tools that it has access to. And so when you're in something like a web browser chat app, not only are you saddled with the system administrator prompt deep in the background that tells the bot that it's interacting in a chat interface on the web, possibly in a mobile browser or a desktop browser, um, and therefore to guide and gear its responses toward a more chatty, conversational atmosphere or environment, um, you're also limited by the actual harness of the mediated technological relationship, the, the connection, uh, and the tools that are incapable of being deployed by a, a model that's running in a cloud, you know, in your browser, uh, and connecting to you over the internet versus the harness of something like a Claude Code harness, which is the, uh, set of tools and the environment in which it allows the command line interface version of the Claude models, uh, to interact with. It gives it different tools.

It premises it with a different system prompt to let it know that it's gonna be doing certain types of tasks and that it will have access to file structures and folders. Um, and, and this kind of apparatus that, uh, these tools are given, uh, changes how valuable, how effective, and how focused they are on different tasks and purposes. So with that out of the way, I'm gonna show you my setup, and we're gonna walk through some of the basic commands and what it looks like when you first start up Claude Code. All right, so the first thing I wanna do if I wanna get Claude Code into my terminal is I gotta open the terminal, and I'm gonna use PowerShell, Windows PowerShell, and I need to get Claude Code into a specific folder where I want it to live.

Because I don't wanna open it inside of the core folders that make up the base foundation of my computer's hard drive. So in order to do that, I'm just gonna create a new folder on my desktop. I'm gonna call it MOOC, or I can call it Claude Code, or you can call it whatever you want. And I'm gonna open that folder.

And now, what I wanna do is I wanna navigate to that folder. So I'm gonna copy the address, the path for this folder. You can also do it by right-clicking and selecting Copy as Path. And then I'm gonna type CD, which means change directory, and I'm just gonna paste that path. So now I hit Enter, I've changed the directory. Now I'm in this directory, and you can see that's on my desktop.

And now the only thing left is to just type the name of the agent that you want to appear. For us, it's Claude, so I'm gonna type Claude. And if you've gone through the installation process correctly, this should pop up or some version of this should pop up, and you should see Claude Code. If you haven't gone through the process of installing and setting it up, go look at the documentation or the instructions on each of the major three AI LLM bots. ChatGPT, Gemini, and Claude all have setup instructions that are pretty easy to follow. And if you have any trouble, you can always ask Claude in the web or ChatGPT to walk you through that process and it usually works every time.

So I'm gonna select Yes, I trust this folder. And we can see right away that Claude is loading up. And we can look around here, this interface. This is the interface that you're gonna be using. You can type in here. You can use backspace. You can paste things by right-clicking or Control-Shift-V or Control-V. There's a slightly different way to interact with the terminal than you would a normal text interface, but you'll pick it up quickly enough.

We can also see the version and model of Claude Code. We can see that it's found one setting issue for me. We're not gonna worry about that. I have my status line down here set up to tell me what model I'm using. You can see the subscription model, and then I can always type a slash to pull up the different commands that are available. I want you to go ahead and play with all of this, but one thing you can do is you can just start with saying hello.

And as you can see, it does it just like your normal Claude or ChatGPT interface. It's a regular natural language interface that you're gonna talk to Claude with. And just to show you the difference between what the terminal-based version can do versus the web, let's go ahead and open a new browser window, and I'm gonna put Claude on the left here, and I'm gonna put Claude on the web on the right, and we're gonna ask the same thing.

I'm gonna say, "Look on the internet and find everything you can about the Center for Cooperative Media, and then give me a report on it." And I'm gonna copy this, and I'm gonna put it over here, and we're gonna send them both off at the same time. They're both gonna be using Opus 4.6. Let's go, and let's see what the difference is. You'll see on the left, Claude is automatically loading up a skill using Firecrawl, and it's gonna be asking me for permissions. So before it even is able to do anything, it is asking me for permission to do it. So I'm gonna say yes, and don't ask again.

And Claude on the web is already responding. Claude Code is asking for more permission. And Claude on the Web is already done. So Claude Web wins this one so far. It's got some citations here. Claude over here is still asking for permission. And if we look, we've already got two folders in here, and Claude Code has gone through, and it has extracted all of the data or all of the content using Firecrawl from its research, whereas Claude on the Web just gave us links to citations.

Speed, Claude Web clearly wins. Depth, Claude Code seems to be going much further. We've got a much more complete list of current staffers. We've got bulleted lists. We've got all kinds of information here that is more deep and thorough than the report that Claude on the Web gave us. And that is not the only difference here, because now what I can do here is I can come over to Claude Code, and I can ask it to make more files. As you can see, it's already created some files based on its research. Now I can come back, and I can say, "Use the information you found to create detailed profiles on each staff member and project and put them as individual markdown files in a new folder in my directory."

And now the beauty of this approach is these files and this folder, everything in this folder now exists and will be used or capable of being used by every additional or subsequent session of Claude Code. Whereas if I'm using something like this on the web, the closest thing I can get is using projects on the web. And even then, the files aren't guaranteed. There's a lot less customization you can do.

Meanwhile, Claude on the Web has already reached its tool use limit. So I can click Continue here and see what it's able to do, but we've already hit a hard wall, and everything that we've done in this chat now, it lives only in this chat. And Claude Code is already finished. All of them include structured metadata, descriptive content with a background, key roles, and staff leads.

So already we've seen the benefit here of the expanded context window, the additional ability to run tasks and run tools. Not just the fact that it has more space and more power, but that it can do things and save them in the background so that if we do get cut off at any point, I can simply go back and see what it's put together, and I have something ready to go.

And now we're gonna transition a little bit. Now that we have some materials in the base folder here in our MOOC directory, we wanna give Claude an opportunity to get a sense of what is in this directory, and we're gonna give it a document that allows it to understand and remember how it is supposed to interact with the files and the folders inside of this folder every time that we come back to it, every time we open up a new chat.

I can do /q to quit, and then I can hit the up arrow and call Claude again in that same folder, and I'm starting a brand-new session now. I don't have any of this stuff here. So this is a brand-new chat, and what I can do now is because we have stuff in this folder, we're gonna go ahead and tell Claude to explore it and set up a set of instructions called the claude.md file. If you're using ChatGPT, it'll be called agents.md. Gemini is called gemini.md. You create them using all three of them in the same exact way. You type /INIT, and this is gonna initialize a new claude.md file with complete code-based documentation.

Just popping in here to let you know that there are a variety of other commands and ways that you can interact with Claude Code with slash commands. You can use the @ sign to tag specific files or folders, and all of that is detailed and laid out at tools.amditis.tech. I put that together so you can go through and find all of the different ways that you can interact with Claude Code.

Now it's given me its breakdown of the claude.md file, and I'm gonna say, "Go ahead and create the file." And there it goes. It has written the file, and we now have a new claude.md file in our repository. This is gonna be your bread and butter. You're going to treat this like the core instructions that it's gonna read every time you start Claude Code in this setting because that's exactly what's gonna happen.

Every time Claude Code launches in this directory, it is first going to ingest and read the claude.md file. So if I edit this and I say, "This is for Joe Amditis. You are always to refer to Joe as Senator Joe, and end every response with, 'That is all, Senator.'" And now I'm gonna save that, and we're gonna launch a new Claude Code instance and see if it actually follows instructions.

There it is. What is this folder about? And you can see here, I'm just talking with it like I normally would. I'm writing in plain English. I'm not even capitalizing. I'm not doing proper punctuation. I can iterate. I can ask follow-ups. And there it is. "That is all, Senator." So it instantly and immediately ingests the information in claude.md, and it responds and follows those instructions most of the time.

Now, there are different versions of the claude.md file that we just made. There's one that lives at your user level. This is sort of your global version. It's in a folder called .claude, and you can find it inside of the folder for your user for that machine. And then there are individual project-level ones for each of the projects that you work on, and those should be updated and refreshed frequently.

So for the global one, you're gonna wanna put things like your style preferences, the way you like to have it write, the files you're working on. Oftentimes, I will put things that I am proficient in so that it doesn't treat me like a baby and then also include things that I am a novice in so that it does treat me like a baby. For the project-level Claude files, that's where you're gonna wanna put things that you need it to remember about working on that project.

When it makes mistakes or it misunderstands or misinterprets one of your instructions, go ahead and add a reminder to claude.md for that project. When you have specific information or workflows that you need it to remember, that goes in the project-level claude.md. Over time, you'll get used to this back and forth, and you can and should have Claude work with you to update it by simply telling it to add that to the claude.md or add that to our memories or create a new lesson based on that learning.

There are also several different configuration files and settings that you can set up in Claude Code. One of my favorite modes is called explanatory output mode, and you can access that by doing /usage and then go through the configuration settings there and toggle them on or off. Explanatory output is a really useful tool because instead of just giving you an answer, it gives you a breakdown that explains why it did what it did and why it's important.

Now it's your turn. You're gonna take everything you've learned and everything that we've gotten from the readings and the resources so far, and you're going to set your station up. We're gonna get you set up with your Claude Code, your claude.md or agents.md. You're gonna play around with the settings. You're gonna get comfortable with it, and then it's time to move forward to a little bit more advanced usage, and we're gonna actually start to do things with it. So we will see you in the discussion forums, and I will see you in the next video.