In this 13‑minute tutorial Isaac from Upstart.co outlines the essential 20% of knowledge needed to capture 80% of Claude’s value. He starts by stressing the importance of having a clear use‑case before opening Claude. He then breaks down Claude’s three interaction modes—Chat (conversational), Cowork (agentic task‑oriented), and Code (developer‑focused)—and demonstrates the difference between chat and task workflows, highlighting the Composer for parallel task handling, project/folder organization, and status tracking. Isaac reviews the available model tiers—Haiku (fast, low‑usage), Sonnet (balanced), Opus (full‑power default), with a brief mention of Fable—explaining when each is appropriate. He introduces five core capabilities: Connectors (pre‑built integrations like Airtable, Gmail, Google Drive), Projects (static context workspaces), Skills (playbooks of instructions, files, scripts for repeatable outputs), Artifacts (plain vs. live, with examples such as tax calculators and KPI cheat sheets), and Schedules/Routines (automated background jobs and event‑triggered flows). Throughout, he provides concrete examples—email triage, audit guideline projects, accounting efficiency audits, daily meeting prep, and a client‑onboarding routine—to illustrate how to pair a chosen use‑case with the right mode, model, and capability. The video concludes by urging viewers to consider these three dimensions for every automation they plan with Claude.
Chapters
Hey, this is Isaac from Upstart.co, and I want to share with you the 20% of what you need to know to master 80% of Claude. We're going to be talking about 3 actions that you need to take, one being mapping what's worth it, finding the use case in the first place that you'll go into Claude with. Don't go into Claude with nothing in your mind of what you want to do. Then we're going to talk about mastering Claude. What are the models, the modes, and the capabilities you need to know? And understand how they differ. And then we'll see some examples of how, how can you pair your use case or certain use cases with Claude's capabilities to know how to automate that use case you found. Let's get right into it. This video in particular, I'll focus primarily on master and match, assuming that you already have a use case in mind. And I'll talk about some examples. So if you don't, at least you can think of one as I walk you through this. Now, what Claude actually is, At its very core, you have different modes where you can use it, different models, how powerful the engine underneath is, and you have certain capabilities that they've added over time that empower those, that engine and that mode where you're working on to do the job. Let's talk about modes real quick. At its core, chat, think about it where you think, co-work where you do, and code where you ship or build. At its core, it's kind of what it is. Chat is primarily a way for you to interact with it conversationally, think, draft, and explore ideas. Cowork is an agent hardness, so it's a little bit more powerful but still user-friendly. You can work with it on your desktop computer with local files. It can run multi-step tasks and think through and plan and execute its plan. And every time you want something done, an actual task with an action item is what I would reach for. Lastly, built for developers initially, but for builders, and don't get intimidated, some non-technical users are also leveraging PlotCode. It's a more powerful version of the agentic work with some unique features to build websites, to build apps, to integrate with GitHub, which is a way to store and host code, and other things that will allow it to be very, very powerful, but not really the focus of today. We're going to focus on chat and code work specifically. Now let's demonstrate these different Claude code modes. First by looking at task versus chat. What's the difference? In chat mode, you'll notice that you have voice mode available. Again, you're having a conversation and almost on every commute that I go on my own, I pull up my phone, I put up my voice mode in Claude and think through a problem I'm thinking about at the moment. On my way home right now, I was thinking about the script for this video. Sometimes I'm thinking about the specific problem that I have in the business and how to allocate our marketing budget. And these things can be helpful to talk throughout loud with a sounding board. Subtle, but you notice that when I'm on chat here, it says new chat. When I move to cowork, it says new task and enables a composer. That's another area that I wanted to briefly showcase. What is a composer? Composer, it's really something that enables you to multitask. By opening a new composer, I can say here, help me draft an email for this client to let them know that we're leaving them or letting them go. And while you run that, you can spin up another one and say, help me draft a presentation for a cold 101 workshop. And you can see these things starting to work in parallel as you work through. And you can see them being worked on here with this little icon that represents this is a task that's currently running. Once it needs some input from me, it will show up as a waiting answer. And once I am done, I'm able to actually, let me show you a quick example, mark it as read or unread. So I have yet to work on it or not. And I can then see what are the active tasks I have on my plate here on Cowork. To not lose track of the things you got started, which can be easily lost in chat mode. You also have the ability to pick and select a project or a folder. Folder primarily when you're working on with your local files and project primarily when you're working with a version of that, but you want to share with your team or you want to live in the cloud. And last but not least, as I mentioned earlier, every time you work with CodeWork, it will put together a progress and a plan, and you'll see that plan in action. Let's see, show you an example from a prior one. Well, you can see the progress is made, and you can even suggest what changes you want it to do for the plan it has. It also shows you connectors and also the history of all the tool calls it did with that particular connector. So it's a lot more robust, as you can see on this sidebar. While On specifically the chat, you really only see artifacts and content for the most part, which are files you've given it and artifacts that it has created, which we'll talk about in one of the capabilities we have coming up next. Briefly looking at code, again, we're going to dive into details, but in code you currently have artifacts, you don't have projects available, and you have routines, which is— we'll talk about later— but a way in which you can get tasks to run on autopilot in the background. Now you have 3 models. You might have noticed that every time I pull this up, I have the ability to choose Haiku, Sonic, Sonnet or Opus. There's also Fable available, but we'll kind of leave that outside of the scope of this conversation. Haiku is when you want something super fast. I barely reach for it, and it's mostly to save usage when I'm near my limits. But if it's something where it has to draft, it doesn't, it doesn't need to interact with any tool or any like major skill, I rely on Haiku for that. Sonnet is a solid middle ground. It's a little bit faster than Opus, which thinks about things for a long time. So if I know it's just one tool that I'm going to be interacting with and it's just creating something, doesn't need to pull a lot of information to make sense of it, it doesn't need a lot of thinking power, I'll move forward with Sonnet. But honestly, 80 to 90% of the time, I'm just opting in for Opus as my default and really only moving down if I'm trying to save usage or if I really want a speedy and quick answer. Here are the 5 capabilities you should know about Claude: connectors that allow you to plug into apps you already use, projects, a workspace where you hold stable and fairly static context, skills where you have saved Playbooks, these could be files, instructions, and scripts to do a repeatable task with a consistent output. Artifacts, which are self-contained mini applications that can be static in nature or pull data from your connectors. And lastly, schedule and routines, which you briefly talked about. Talking about connectors, connectors is where you plug Claude into the apps you already use. And if you're using the desktop app or the browser version, you can go to customize, in the connector section to see what are some of the apps you have available. These are apps I already have connected like Airtable, Calendly, ClickUp, Fathom, Gmail, Google Calendar, Google Drive, and more. And you can click on add to explore some of the native connectors Claude has available as well as adding a custom connector if needed. If you wanted to answer something like what emails are waiting on my reply this week and it needs to pull from Gmail, from Outlook to know that, that's how it'd be able to get that information. So think about what tools do you have connected to Claude today and which one you benefit from connecting in the first place. Projects are a dedicated workplace that holds context and files and instructions. We would really like to see it and only use it when the context is fairly static. And an example we did for this, we're working with an audit firm that has some guidelines to review statements, and that guidance only changes once a year. It was fairly a good use of projects to have the files available in the context of that particular project, the instructions of how to review, and that the team can come here to do the reviews easily. It's not my favorite feature, but think for yourself if there's a client, a project, a type of task like the reviews we talked about that you do regularly that have a fairly static context where you can spin up a project to get it done. By the way, projects you can share with your team. As well. Skills is a capability that you can use to have a safe playbook in the form of instructions, files, and scripts to be able to do and get a consistent and identical output every time. It's one of those most highly recommended capabilities to automate work within accounting firms. An example, we have internally something we call an efficiency audit where we assess what are the biggest opportunity a firm has. And when we present a roadmap with a list of options, we do use a skill that enable us to categorize, estimate, and have an Excel layout on how that should be done, as well as script that's able to create it to begin with. So think for yourself, what's something you would benefit from turning into a skill where an example file, a script of what the output looks like, and some instructions so you get a consistent output would be beneficial? Anything that's reconciliation, working with work papers, pulling data into an Excel to then reconcile and tie out balances, are great examples of that. Let's talk about artifacts. You have 2 types of artifacts, a plain artifact available on chat and code and a live artifact available on co-work and code. And the only difference being that you're able to pull from the connectors you have available. For example, if you have a client tax estimate calculator that takes some inputs, gives you an estimated liability, but doesn't need to store that data long-term, artifacts is a perfect use case. for that. Live artifacts, however, if you want to pull from your practice management system, what are some open items? If you want to pull from your inbox, what are some unread emails? That's where you need to have that connector available and use a live artifact instead. Some examples of artifacts are the presentation you're looking at right now is an artifact. As an example, we sometimes have answering a particular question for a client and we might build like a simple— this is an example of a KPI cheat sheet that we built. This is almost like a brochure of some sort, a simple web page if you like. And on the live artifact, as we mentioned, if we wanted to pull from our email on a regular basis, we'd need to use something like live artifact connected to our Gmail connector or Outlook connector to be able to parse apart what emails we've gotten in the last 7 days that are unread and get it to categorize them as high priority, not yet replied, or noise. Last but not least, if you want an automated execution, you want something to run routinely on a regular basis or based on a particular event, for example, where a task is completed or when an email is received, That's where you use scheduled and routines. Scheduled is available within Cowork where you'll be able to, as an example, have a daily meeting prep, have a daily inbox triage, have a way to on a weekly basis prioritize what should you be doing as a firm owner. And to showcase the daily meeting prep as an example, you might have here the outputs is a document that tells me, hey, you have a check-in with this team member. These are the things you talked about earlier that you might want to reference again. You have a follow-up with this particular potential client. This is when you promised and you talked about what their priorities were and what you wanna do moving forward. You have this AI accelerator that you want to go to. It conflicts with your thing, but perhaps you can join at this time. Like, it looks at every single meeting I have for that day and helps me prepare for it with a simple 3 bullet points that reference the things we've talked about before or meetings I've had before with that particular person. Like the cohort sidebar that we share, it has progress, working folders, and context. The only difference is that now you also have runs because this is supposed to run on a regular basis. Every time you have something like this that you wanted to run without relying on your computer, that's where it makes sense to move into cloud code and create a routine instead. Because in a routine, you'll be able to get this to set on a schedule and run on cloud servers rather than on your computer. This is where you also be able to trigger it from a specific event. For example, we have a registration form where people have to put in a deposit to lock in their spot to work with our agency. And once they do that, it triggers a series of things. It adds them to our client tracker, pulls meetings we had, research that person, pulls everything into the task description almost as a handoff document to the delivery team. It creates a new folder on Google Drive. It announces to the team that now we have this new client available, and it drafts a welcome email that I just have to check, adjust, and send through. So ultimately, every time you want to automate something in the background, reach for schedules or routines in Cowork and Code respectively. There are some other features like your ability to steer Claude halfway through a task, which we can do. Within code, now chat, having local file access and local connectors, same thing, as well as browser and computer use, but those are a little bit more advanced and outside the scope of today. And last but not least, it's worth you thinking for every use case that you have, what is the mode you would most benefit from, the model, and the capabilities you need to make that automate that process within Claude itself. That's 20% of what you need to know about Claude to get 80% of the benefit. Stay tuned for a coming video to talk about use cases and how to pair use cases with Claude features coming up next.
