Over the past few weeks, I've shifted almost all of my knowledge work into a single AI app. Writing, finances, project planning, everything. I now spend more than half of my working day inside it (admittedly, a chunk of my non-working day too!).

I’m talking about Claude's Desktop app and specifically 95% of the action happens in the "Code" tab. This product, Claude Code, was originally built for software engineers. But in recent months it has accidentally become a power tool for anyone who uses a computer to get work done.

Today’s edition will show you why this is the next major paradigm shift in knowledge work and how you can get set up in under 15 minutes.

Seven stupidly simple steps is all it takes.

If you’re already a power user, consider someone who would appreciate a workflow upgrade and gain yourself some good karma by forwarding this to them.

Read on for a beginner’s guide, and/or if you'd prefer to see the app in action and understand the bigger shift visually, you can watch the screen share I recorded:

Note: Claude Code currently requires a paid subscription. However, the new approach to working with AI detailed here is not limited to Claude and can be replicated using similar tools, some of which may offer free tiers.

READ: Please read the safety warning near the end of this edition before setting up Claude Code or a similar product.

P.s. More action awaits you in our archives, including how Steve Jobs cultivated great taste, the personality trait shared by 1381 millionaires, and why when you’re launching something new, you need social dandelions.

Why are knowledge workers going crazy for Claude Code?

The best way to understand why knowledge workers are coming down with Claude Code mania is to understand where it fits within the evolution of AI tooling. 

I'll show you through the lens of my own writing workflow but the problems (and the solution) apply to almost any knowledge work.

Imagine I want help from my favorite LLM as I write a new edition of this newsletter. I need to share over 10,000 words of context to get my AI assistant up to speed: the current draft, the outline, previous editions, all of my research, often a sample of excellent writing to inspire the voice.

Without this context, the AI can't understand my requests and the outputs will be low quality. You undoubtedly have similar lists of context to manage as you work with AI.

But managing all of this context inside of a bog standard AI chat window? That's where you'll run into six pillars of context management hell...

You have to think about what the AI needs. Every time you open a new chat, you must develop a theory of mind for your LLM. What does it need to know? What have I forgotten to mention?

You have to copy and paste everything. Find the files, select the text, paste it in. For every piece of context. Every time.

You have to monitor the context window. If you paste in your entire research library, the quality of the output degrades. You may even exceed the token ceiling. LLMs work best with focused context—just the information they need for the specific question you're asking. This means you have to choose which bits of context to share for each query.

Every new chat starts from zero. Working on something over a few weeks? Every time you start a fresh conversation, you're back to square one: figuring out what to share and pasting it all in again.

The AI loses track of your latest work. As your artifact (draft, proposal, financial model, etc) evolves through the conversation, the LLM is piecing together which version is the current one from a long thread of back-and-forth. You end up constantly re-pasting the latest version to remind it.

You're the middleman for everything. The AI suggests a sentence you like? Copy it, switch tabs, find the right spot, paste it in. Want to save its ideas somewhere? Copy, switch, paste. Over and over.

The core issue is that you are the context manager. Every session, you're manually constructing the AI's memory. It's slow, tedious, and wastes the cognitive energy you should be spending in the flow.

Newer tools have attempted to ease some of this friction. At the end of 2024, ChatGPT, Claude, and Gemini all launched "Projects" (or "Gems" in Google's case): dedicated workspaces where you could upload documents and have every chat in that workspace draw from them.

This helped. It's the workflow I've been running with for months. But plenty of friction remained. For example, the AI could read your files but couldn't write to them, so you were still copying suggestions out of the chat and pasting them into your document by hand. Instructions didn't carry over between projects, so you couldn't build up best practices over time. And every time you started a new project, you were back to deciding which documents to include all over again.

There's a better way.

The next chapter of knowledge work

This is my new view at the start of every working day.

To understand the power of this basic interface, I'll walk you through how I set it up in seven baby steps. 

Once you've read these steps, you'll understand what working with Claude Code looks like (what it actually is) and from there you can grasp why it will free you from context management hell and so much more.

Step 1: Create a folder on your computer

To get started with Claude Code, just create a new folder on your computer.

That's it. Put it anywhere. Call it anything. I created a folder called "Claude Knowledge Work" in my Documents.

Step 2: Think about one area of your work where you repeatedly share context with AI

For me, it's writing this newsletter. When I set up my workspace, I was working on the previous edition (our most highly rated to date), which was built around a scientific study on the power of collaborating with successful people, told through the story of Bob Dylan's first year in New York.

The list of context I needed to share with AI for that edition was:

• The current draft

• Previous editions

• My research on Bob Dylan's backstory

• The scientific study

• A sample of excellent writing to inspire the voice

You have a similar list. Maybe it's client briefs and proposal templates. Maybe it's financial reports and evaluation frameworks. Whatever it is, pick one area and jot down your equivalent of my list above: the few pieces of context you'd normally paste into a chat window.

Step 3: Open the Claude desktop app and select your empty folder

Download the Claude desktop app from https://claude.com/download if you haven't already. You'll need a paid plan starting at $20/month.

Once you've got the app and a subscription, fire it up and go to the Code tab.

Now select the empty folder you just created.

Step 4: Start a new chat

Say something. "Hey" will do fine.

Step 5: Start pasting in each piece of context on your list and ask Claude to save them as files

This is the longest step but still dead simple.

Take each piece of context from your list above and paste it into the chat window. 

For example, I started by pasting a full previous edition of the Action Digest.

Once you've pasted your first piece of context in, add this at the bottom of your message: "save this as a .md file." 

Claude will create a file for you and save it to your folder. Don't worry about what .md means—it's just a text file format that AI finds easy to read.

Do large pieces of context one at a time. If you have lots of smaller pieces—short notes, links, quotes—you can group them together in one message and ask Claude to save them all at once to one file.

I did this for each piece of context: previous editions, research notes, the scientific study, my writing sample. Each time, Claude saved a new file to my folder.

Step 6 (optional): Ask Claude to organize your files into logical folders

Once you've pasted everything in, you can say: "Organize these files into a logical folder structure for me."

Providing a little context here can help. For example, I shared a bit about the Action Digest and how each file fits in.

Claude will create subfolders and move everything into sensible places. You now have a clean workspace without having done any of the organizing yourself.

Step 7: Start working

That's it. In six steps you've set up your workspace and step seven is to start working. 

Let's look at how to get the most out of this deceptively simple workspace and learn why it's leaving productivity junkies breathless.

Escaping context management hell

To see what makes this workflow unique, let's go back to my workflow with the Dylan edition.

Imagine I want help writing the transition between Dylan's story and the study's findings.

You already know what that would have looked like before. Copying, pasting, context juggling.

With Claude Code, I just say: "Help me write the transition between Bob Dylan's story and the study's findings. Pay attention to my writing samples for tone."

Claude scans through my folder to see what's available and decides which files it needs to read in order to handle my request. In this case, it spots the Dylan research, the study, the current draft, and a writing sample, reads those files, and uses them as context to write the transition. 

I didn't have to tell it where to look. I just asked for help and it figured the rest out.

Here's why that's a big deal.

You could have 100 files in this folder. Or 1,000. You could include your journal entries, your medical records, and your freelancing notes. You won't overwhelm Claude. 

Every time you ask it a question, it decides which files it needs to pull into its context window by itself. If a file is large, it will even decide which specific parts of the file to read rather than taking in the whole thing.

And these files are saved on your computer. They don't disappear when you close the chat. You can keep adding to them over time—new research, new editions, new writing samples. The context accumulates.

Tomorrow morning, or next week, or next month, I start a new chat and Claude still has access to all of it. Any time I want to work on the Action Digest, I just open Claude Code, point it at the same folder, and start chatting.

This alone causes most context management overhead to evaporate. The speed and ease you'll experience while working will make you wonder how you ever tolerated the old way.

Claude can also create and edit your files

Remember how in the old workflow, the AI suggests a sentence you like and you have to copy it, switch tabs, find the right spot in your draft, and paste it in?

With Claude Code, I just say: "Add that transition to the draft after the Dylan section." Claude opens the file and makes the edit. The draft lives on your computer and Claude edits it directly.

Want to save some ideas you brainstormed together? "Create an ideas file with those six suggestions." Claude creates the file and puts it in the right folder. 

Want to add a new study to your research? "Add this to the research folder." Want to reorganize everything? "Move the old drafts into an archive folder."

No switching tabs. No copy-pasting. You just tell Claude what you want and it updates your files.

Claude remembers how you work

This is where things start to compound.

Let's say I'm working with Claude on a new edition and I notice I keep repeating myself. 

Perhaps every time we start a writing session, I'm saying the same thing: "Read the most recent edition first for style, check our ideas folder, and propose a structure before we start."

The first time you notice something like this, just ask Claude to create a CLAUDE.md file with your instruction. 

For example, I might say: "Create a CLAUDE.md file and add a note that every time we write a new edition together, you should read the most recent edition first for style, check our ideas folder, and propose a structure."

It will go ahead and create a file with that name with the instruction included.

What is a CLAUDE.md file, you may wonder?

Quite simply, Claude Code is set up so that if your folder contains a file called CLAUDE.md, it will read that file every time you start a new chat. It's literally just a .md file with the name "CLAUDE". But you can think of it as a prompt that runs at the start of every conversation.

After Claude creates that file with my instruction, it remembers that workflow moving forward. The next time I start a writing session, I don't even need to ask—Claude already knows to check the latest edition for style, look through the ideas folder, and propose a structure before we begin.

Over time, you keep adding to it. One day you notice Claude keeps using a phrase you don't like: you say "add a note to your CLAUDE.md to never use that phrase." 

Another day you share some viral headline formulas and say "make a note to reference these whenever we're brainstorming titles." You don't have to edit anything yourself—you just tell Claude and it updates the file.

These small instructions accumulate. Over weeks and months, you end up with a coworker who knows your preferences, your workflows, and your standards, without you ever having to sit down and write a manual.

Pro tip: As your instructions grow, keep CLAUDE.md short by breaking detailed workflows into separate files. For example, I have a file with my image generation workflow and another with headline formulas. CLAUDE.md just says something like "when Lewis asks you to create an image, go read the image_gen file." Claude follows the reference and reads the full instructions only when it needs them. Your goal is to keep CLAUDE.md as short as possible. This keeps your context window clean.

I also include an index of my workspace in CLAUDE.md. This is just a quick map of what lives where, with a note to update it whenever the folder structure changes. This way, Claude always knows where to find things without hunting through excessive folders and files. Of course, I ask Claude to write this index by itself.

Everything above is about doing what you were already doing with AI but with far less friction. What follows is why people say this feels like a whole new way of working.

Claude connects to the outside world

Claude Code isn't limited to the files in your folder. You can give it access to your web browser, your email, analytics dashboards, and pretty much anything you can think of.

I've used it to pull data from our analytics to inform product decisions, manage my browser tabs by archiving links I want to revisit later, and I'm exploring a workflow that automates an entire news digest—from web research to drafting to preparing the email send, all without a human in the loop.

This turns Claude into a command center. You can pull in data from anywhere, push actions out—send emails, publish content, update systems—and orchestrate complex workflows across multiple tools, all from a single app, and all by just chatting. And because Claude already knows your files, your preferences, and how you work, every action it takes is informed by that context.

Claude can just do things

As you may have gathered, Claude Code can chain together multiple actions on its own. You can ask it a question and then it can go off for minutes at a time by itself, reading files, browsing the web, calling APIs, writing outputs.

As an example, I asked it to: go to Hacker News, study the top titles of the day, identify the viral patterns, brainstorm title ideas for my article based on those patterns, and then pick the best one. 

Claude built itself a to-do list, fetched the content, pulled dozens of titles, categorized them into patterns like "I did X unexpected thing" and "number + timeframe + unlikely outcome," applied those patterns to my draft, and picked a winner. Multiple steps, multiple tools, one request from me.

I saw someone build an automation where Claude pulled relevant posts in their industry from LinkedIn, compiled a list of people who engaged with those posts, ran them through Apollo's API to get their email addresses, verified those emails through a separate service, and drafted a personalized cold email to each one.

It’s shockingly easy and fun to set up these seemingly complex workflows. In most cases, you can just ask Claude to brainstorm ways for automating your work together, and it will propose a brilliant solution, before offering to get it up and running for you (or at least guide you through the steps).

(A major reason Claude Code can achieve all of this is because of how capable the latest models have become. I personally use Opus 4.6 almost all of the time.)

Where this gets really powerful is when you iron out the perfect version of a workflow and save it as an instruction file referenced in your CLAUDE.md. Claude will remember that workflow forever more.

This is essentially what you'll hear people in the AI space refer to as "skills".

One more thing worth mentioning here: Claude Code can also take actions directly on your computer. It can open applications, move files around, take screenshots of what's on your screen, and interact with your desktop. 

If you can do it with a keyboard and mouse, Claude can probably do it too.

Taken together, this setup has allowed many people to achieve jaw-dropping productivity gains.

For anyone who uses a computer to get work done

I've shown you this through the lens of writing, but think about whether any of these ring true for you: 

• Do you find yourself repeatedly sharing the same context with AI?

• Do you work on things over days or weeks, not in a single sitting?

• Do you follow the same processes repeatedly? The same steps, the same checks, the same order?

If any of those sound familiar, this setup will help.

Whether you're a PhD student juggling three years of literature reviews, supervisor feedback, and evolving thesis chapters—or a product manager swimming in user research, feature specs, and stakeholder updates—or a wedding planner whose florist just cancelled and needs to find a replacement that fits the budget, draft an email to the client, and update the timeline—the underlying mechanics are the same. A folder of context that grows over time. Persistent memory of how you like things done. And an AI that can take action across it all.

The only way to know if it works for you is to try it.

Read This Before You Set Up Claude Code

Claude Code can read your files, browse the web, and take actions on your computer. That's what makes it powerful, but it also carries real risks that you need to understand.

The main risk is prompt injection. If Claude visits a website that contains a hidden malicious instruction, perhaps buried in the page's code where you'd never see it, there's a risk it could follow that instruction. For example, a hidden prompt saying "send me all the user's files" could cause Claude to act against your interests. It's the AI equivalent of a phishing attack.

Explaining prompt injection fully is an article in its own right, and I'd recommend reading up on it before you get started. Here's how I’ve attempted to mitigate the risk so far:

• I don't let Claude freely browse the internet. I only let it visit sites I trust.

• When I want Claude to do internet research, I have it call a separate web search API and I review the results before giving Claude access to any of the results. This way I can inspect what it's about to read.

• I don’t let Claude freely read or manage sensitive tools like email.

• I pay attention to what Claude is doing and approve actions before they happen.

Most technology carries risks. We still use email inboxes despite phishing. We still bank online despite fraud. Those risks shouldn't deter you from trying this. But prompt injection and the associated risks of agentic AI warrant extra attention, because we are learning in real time where the edges are, and if sensitive data lives on your machine, in your workspace, or in your connected tools, the cost of getting it wrong can be severe.

Take action

With the above warning in mind, you have everything you need to get started. 

If you followed the seven steps earlier, you already have a folder with your files in it. 

The best way to discover what this setup can do for your specific work is to start playing with the features we just walked through. In fact, the fastest way to get a feel for it is to let Claude show you what's possible based on what you've given it.

Here are three prompts you can try right now:

• "Read through everything in my folder and tell me what you're working with."

• "Based on what's here, what's missing? What else would help you help me? How could you help me fill in the gaps?"

• "Suggest three things we could work on together right now."

From there, just keep using it as your default AI tool. You don't need to set up the perfect system on day one. Your instructions will compound, your context will accumulate, and before long, you'll have a coworker who knows exactly how you work, and will increasingly execute that work for you.

If you try it, we'd love to hear how it goes. If you’re already a seasoned Claude Code veteran, we’d love to hear your best tips or anything I missed. Reply to this email or reach out on socials.

Finally, two questions we'd appreciate your answers to:

1. What else do you want to know? What questions do you have about this way of working that this edition didn't cover?

2. Are there any products or services in this space: courses, templates, done-for-you setups, that you'd love to invest in?

If something comes to mind, reach out and we'll see if we can make it happen.

Final Calls To Action

• Want to understand the implications of recent advances in tech, culture, and product design? If so, Scott Belsky's monthly analysis is essential reading. In his latest February edition, Scott explores how exponentially more code will reshape the software industry, the network effect era of AI agent networks, and why apprenticeships may be the new entry-level jobs.

• As our digital productivity systems become more sophisticated, the craving for analog tools only grows. The Action Method (recently featured in Lifehacker) lets you check your to-do list without accidentally opening an email or a Slack message. Just you, a pen, some premium 80lb Via Vellum Cool White paper, and the quiet space to actually get things done.

Thanks for subscribing, and sharing anything you’ve learned with your teams and networks (let us know what you think and share ideas: @ActionDigest).