Getting more out of Claude
0 of 9 lessons done

There are three Claudes: one you think with, one you delegate to, and one you build with. In this session you take one real task and run it through all three, testing a couple of models along the way, so you learn the differences by feeling them instead of reading a chart. Go at your own speed. Nobody is waiting on you.

How this works

Three sections that build on each other: think with it (Claude Chat), then delegate to it (Claude Cowork), then build with it (Claude Code). It's the same skill each time, pointed at more reach.

What ties it together: in the first lesson you pick one real task, and you carry that same task through all three Claudes. By the end you'll have done it three ways and you'll know, for your own work, which Claude and which model to reach for.

Bring a task you already have, or pick one from the menu in lesson one. Both work.

Think with it: Claude Chat

The Claude you already have open. You ask, it answers, you're driving the whole time. This is where everyone starts, and where most of the money-saving habits live. Pick your one task here.

Chat 1Meet the three Claudes, and pick your one task

People say "Claude" like it's one thing. It's really three, and they differ by how much you hand off:

The shift to hold onto: you're not learning three separate tools. You're learning one skill, talking to a capable coworker, and pointing it at more and more reach. Same skill, bigger job.

The best way to feel the difference is to do the very same task in all three and notice what changes. So before anything else, pick that task. Choose something real and low-stakes with a few steps in it, not a one-line question.

Pick the one task you'll carry through
Mark complete when: you've named your one task in a sentence, and you can say in plain words what each of the three Claudes is for.
Chat 2Do it in Chat, and match the model to the job

Start where you already are. Open Claude Chat on your laptop or phone and do your task. You ask, it answers, you steer. Nothing new yet, and that's fine.

Here's the part most people never touch: inside any Claude, you pick a model. Heavier means smarter, but also slower and pricier. Most work does not need the heavy one.

The shift to hold onto: don't send Opus to reformat your grocery list. Start at Sonnet, and go bigger only when it's actually stuck.
Do this
  1. Do your task on Sonnet (the default). Get it to something usable.
  2. Take one small piece and re-run it on Haiku. Was it good enough? Faster? For a lot of little jobs, it is.
  3. Take the hardest piece and run it on Opus. Did the extra thinking actually change the answer, or not really? Either way, now you have the feel for it.
Watch out for: the model picker is per-chat, and it's easy to forget which one you're on. Glance at it before a big ask.
Mark complete when: you've done your task in Chat and compared the same ask on at least two different models, and you have a gut feel for when the heavier one earns its cost.
Chat 3One clear prompt beats ten vague ones

Most wasted back-and-forth starts the same way: a lazy first ask, then eight rounds of "no, more like this." Claude isn't a mind reader yet. A clear first message does in one pass what ten vague ones do slowly.

Think about ordering a taco. "Make me a taco" leaves Claude guessing what kind. "Three flour tacos with carne asada, onions, cilantro, and salsa verde, all on one plate" gets you the thing you actually wanted. Same idea with any task.

Tell it four things up front: who it's for, how long or how much, what tone, and what "done" looks like. Put those in the first message, every time.
Paste this into Claude
Here's what I need: [your task]. It's for [who it's for]. It should be about [how long or how much]. Tone: [the tone you want]. I'll know it's done when [what "done" looks like]. If anything's unclear, ask me before you start.

Fill it in for your own task and send it. Notice how much closer the first answer lands when you say those four things up front.

Mark complete when: you've turned a vague ask into one clear one and gotten a usable result in a single pass, not ten.
Chat 4Don't let one chat run all day

Claude has no memory between messages, just the chat in front of it. So every time you hit send, it re-reads the whole conversation to reply. It's like skimming the entire group chat before every single text. The longer the chat, the more it re-reads, and the more it costs. Long chats also get worse, not better, as the thread fills with old tangents.

The fix: keep chats short and about one thing. Start a fresh chat when the topic changes, or when one's just gotten long.

"Start a new chat" sounds like losing everything you built up. You don't. You carry the important bits over. Three ways, easiest first:

Paste this before you leave a long chat
Sum up everything we've decided and figured out in this chat, so I can paste it into a fresh chat and pick up right where we left off without losing anything important.
Mark complete when: you've asked for a handoff summary and started a fresh chat with it, or set up a Project that holds your standing context so you stop re-pasting it.
Chat 5Keep your context in files you own

The context note from the last lesson works, but a note buried in one chat gets lost. The upgrade is to keep your standing context in real files you own, in one place, so any Claude can read it. Write it once, reuse it forever.

A context file is just a plain document about you and your work: who you are, what you're working on, how you like things written, and the projects and people that come up over and over. It's the same idea as a Project's memory, except it lives in files you keep, not locked inside one chat.

Where Obsidian comes in: Obsidian is a free app that keeps a folder of plain text notes (it calls the folder a "vault") right on your computer. Nothing fancy. It's a comfortable home for these files, and because they're plain files in a folder, Claude can read them directly in Cowork and Code, or you paste them into Chat and Projects. If you'd rather not install anything, a plain folder of text files works exactly the same way.
Do this
  1. Make a home for your notes. Install Obsidian from obsidian.md and create a vault, or just make one folder somewhere easy to find. Either way, you now have a place your context lives.
  2. Write an about-me file. A few short sections: who I am, what I'm working on right now, how I like things written, and the people or projects that come up a lot. Keep it plain and real.
  3. Add one project file. A second note, like project-name, with what it is, what's decided, what's still open, and the details Claude keeps needing.
  4. Feed it to Claude. In Chat, paste the file at the top of a new chat, or drop it into a Project so every chat already knows. In Cowork or Code, point Claude at the folder and let it read the files itself.
Not sure what to put in it? Paste this into Claude
Interview me to build a short context file about me and my work that I can reuse with you later. Ask me one question at a time, covering who I am, what I'm working on, how I like things written, and the projects and people that come up a lot. When you have enough, write it up as a clean file I can save and reuse.
Watch out for: don't put anything in a context file you wouldn't want an AI to read: passwords, secrets, private financial or medical details. It's for the stuff that helps Claude help you, not a vault for sensitive information.
Coming up: you'll reuse this same context file in the next two sections. When you delegate a task in Cowork and build in Code, you can point them at these files so you never re-explain yourself.
Mark complete when: you've got at least one real context file saved in a folder or an Obsidian vault, and you've used it in a fresh chat, whether you pasted it in, added it to a Project, or pointed Claude at the folder.

Delegate to it: Claude Cowork

Same coworker, more reach. In Chat you drive every step. In Cowork you hand off the whole task and step back while it does the clicking. Do the Chat section first, and bring the same task with you.

Cowork 1Hand the same task to Cowork

In Chat, you were doing the task together, one message at a time. Cowork is where you delegate the whole thing. You describe what you want and what "done" looks like, and it goes off and works through the steps, clicking through your apps and files for you while you watch.

Take the exact same task from lesson one and hand it over as a job, not a step-by-step. This is where the clear-prompt habit from Chat 3 pays off, because now you're briefing a coworker who's going to run with it.

What "handing it off" looks like
Watch out for: Cowork actually does things, so keep it on your low-stakes task and watch the first few steps. If it heads the wrong way, say so and it adjusts, exactly like you would with a coworker.
Mark complete when: Cowork has taken a multi-step version of your task and worked through it while you stepped back, and you watched at least part of it happen.
Cowork 2Notice what changed: Chat vs Cowork

You've now done the same task two ways. This little bit of noticing is where the whole session earns its keep. Open a note and actually write down what was different, while it's fresh.

Write down

If you want, run one piece on a different model here too, and see whether it changes how well the hand-off goes. You're building a real feel for what fits what.

Or let Claude help you compare, paste this
I did the same task two ways: first in plain Claude Chat, then in Cowork where I handed off the whole thing. Here's what happened each time: [describe both]. Help me list what each one was better and worse at, so I know which to reach for next time.
Mark complete when: you've written down at least three things Cowork did that Chat couldn't, and one place it struggled or needed you.

Build with it: Claude Code

The same task, pointed at the most capable Claude. Where Cowork does the task, Code builds a thing that runs and keeps running. This is the section that opens the most doors. Do the first two first.

Code 1Hand a build version to Claude Code

The third Claude writes and fixes actual software. You don't write the code and you don't need to read it. You describe what you want, Claude Code builds it, you look at the result and say what to change. The rhythm is the same one you've used all along: describe, look, adjust.

So take your task one step further and ask Code to turn it into something that runs:

Where to do this: you can use Claude Code right inside the desktop app, or in the terminal for the full experience. Either is fine here. The desktop app is the gentler place to start.
Want the full build path? From Chat to Code walks you all the way through Claude Code, from the desktop app to the terminal to putting a real thing online. This lesson is just a taste of the third Claude.
Mark complete when: Claude Code has made something real for your task, and you've opened or run it.
Code 2Write your own Claude cheat sheet

You've run one real task through all three Claudes and tried a few models along the way. That gives you something no slide can: a felt sense of which one fits your work, and when. Write it down while it's fresh, in your own words.

Don't aim for a perfect chart. Aim for the notes you'd give a friend who does what you do: for my work, reach for Chat when ___, Cowork when ___, Code when ___, and start at Sonnet unless ___.

Paste this to draft it with Claude
I did the same task three ways: Claude Chat, Cowork, and Code, and I tried a couple of different models (like Haiku, Sonnet, and Opus). Here's how each one went: [your quick notes]. The kind of work I do is [describe your work]. Help me write a short "which Claude, which model, when" cheat sheet that's specific to me. Keep it to one page.

Then pin it somewhere you'll see it. That cheat sheet is now your standing context, and it's the kind of thing worth dropping into a Project so every future chat already knows how you like to work.

Mark complete when: you've written your own "which Claude, which model, when" cheat sheet, based on having actually done the same task three ways.

You've met all three Claudes.

One real task, done three ways, with a couple of models tested along the road. You don't have to guess which Claude to open anymore. Go point it at the next thing.