Hey Reader,

As an entrepreneur, you either:

1. Delegate a task
2. Delete a task, or
3. Do it yourself

If you want to delegate it, you need to hire someone or something.

You can hire:

1. Human(s)
2. AI(s)
3. System(s)

In my previous email, I've talked about how the length of tasks an AI can handle is rapidly increasing. But does that allow us to actually hire them and delegate full workflows?

Well, the answer is yes and no.

No, because when you hire a human to do something, you not only delegate the task itself, but also the management overhead of having that task on your to-do list, and the responsibility of the task.

If you delegate the task to an AI (like Claude Cowork) you can only reliably delegate the task itself, and you'll still have to manage the AI to do it right, and you're responsible if things go wrong.

We're still at the stage where you have to micro-manage the AI a little bit.

That's exhausting, so here's what I do so that I can manage my AIs a bit less. It's still not 100%, but it's good enough for me.

Context engineering by design

I currently do most of my AI work in Claude Code and Cowork. If you're not technical, you can use Cowork the same way. I'll show Claude Code screenshots to better explain what's happenning but they're the same.

Claude Cowork is a Claude Code wrapper for non-technical people.

If you know how to start a coding agent inside a folder, like Codex or Antigravity, the concepts work the same.

When you start a chat in Claude Code, it shows a similar statusline (this is contextbricks for those nerds who are interested):

Look at the last line. With 0 messages sent, we've used up 0% of our context window, and have 1000k tokens free (which is the 1M context with Opus 4.7).

Now let's see what happens if I say hello:

It used 40k tokens! WHAT? 40k tokens for 8 words?

You might be wondering how that is possible. I didn't ask it to read any file or anything like that.

It also would have cost me $0.27 in AI compute costs if I was paying per usage, but I have a monthly subscription so that's more of a fun metric to see. If I reach my cap, I'll just get rate limited.

Now let's try a prompt that asks to work in a project. Let's say I keep developing my small-business-plugin for Claude Cowork:

It gave me a long answer I had to collapse so we can fit on an image. It read a few files, listed the directories, read a few more files, checked my work history and then gave me an answer on the project's status.

If we use the /context command, we can see a detailed breakdown of what ate up our context window.

We are now 46.5k tokens in, which breaks down as follows:

• ⛁ System prompt: 8.6k tokens (0.9%) — written by Anthropic
• ⛁ System tools: 9k tokens (0.9%) — configured by Anthropic
• ⛁ Custom agents: 3k tokens (0.3%) — not important
• ⛁ Memory files: 5.3k tokens (0.5%) — memory from previous chats
• ⛁ Skills: 6.4k tokens (0.6%) — only the descriptions of them
• ⛁ Messages: 17.6k tokens (1.8%) — this is the important part
• Free space: 917.2k (91.7%)
• ⛝ Autocompact buffer: 33k tokens (3.3%) — to compact the conversation and keep going.

The Messages part and the Skills part is where the magic happens:

Messages includes everything that I've sent, as well as all files that Claude opened, worked with, etc.

When I asked it to work inside the small-business-plugin-dev project, it opened:

• The main CLAUDE.md file I have in my workspace: gives it context about me, my business, the workspace, the skills, and sets its behaviour to my preferences
• The CLAUDE.md file inside the project folder, which gives it the project specific context, describes the folder's strucutre, the status, etc.
• Any other files it thought was needed to read to comply with my request.

Now the chat is ready for me to work with it.

The AI now has full context on what the project is about, what we're trying to accomplish and what's next.

I have each of my projects set up this way, but I use one main Claude Code & Cowork folder for everything, so the AI can look at other projects if I want it to.

When I start working on a new client project for example, I create a new folder, titled with the client's name and the project, and I put EVERYTHING there:

• meeting transcripts
• the proposal I sent out
• my notes
• the contract
• the deliverables
• everything

It literally only takes 1-2 minutes to set it up:

1. I usually just ask Claude to "setup a new client folder for X project, make a meeting transcripts folder inside".
2. Then I manually export the transcript from Sembly, and put it into that specific folder.
3. Then I send the prompt: "I've put in our meeting transcript, what's your understanding of the project?"
4. Then the steps follow based on whatever the project is. I ask Claude to create files, to create the proposal, to strategize the project, to create a timeline, to create tasks, etc.

At this point I also use the /rename feature to name the conversation, so I can use the /resume command and resume it any time.

One project = one folder = one ongoing thread (until that gets too big)

This also makes context engineering fully automated essentially. All relevant files live somewhere inside my main Claude Code & Cowork folder, and all files are listed in either the project's CLAUDE.md file, in the main CLAUDE.md file, or easy to be listed with a "list files" command by Claude.

So with this, you've setup the ground for your AI employees. They have a desk and a workspace.

Now it's time to teach them the SOPs.

Skills = AI employee SOPs

Claude Skills are great because they make the tasks you use AI for repetitive and consistent.

You should be building Skills for everything in your business. Without skills, you won't be able to delegate tasks to AIs, and won't be able to remove yourself as the bottleneck for the tasks.

I currently have around 40 skills for my business, maybe even more. Some are just a text file SOP, some also include actual tools, scripts and agents that the AI can operate.

So I'm giving the AI employees three things:

1. A safe space to express themselves (the folder)
2. An SOP to follow (the skill)
3. Tools to act in the real world (tools, scripts, MCP, etc.)

I currently work around 30-40 hours a week, don't work after 8 P.M, don't work weekends or holidays, have the highest sleep and recovery scores I've ever had (92%+), workout 3x a week when I want and make enough money to not only put food on my table and sustain my business but also to have fun in my free time.

If this sounds like something you want here's how I can help:

1) I made a 23 minute video on YouTube that's a full-guide for creating Claude Skills. You should watch it and start delegating your tasks:

​

2) I also made a 13 minute video on YouTube on how to setup your Claude Cowork folder for efficient work. Watch it and set it up:

​

3) I have a full course on using Claude and delegating your tasks to it in the Paladin Guild, my AI implementation community for non-technical founders. It's currently closed for enrollment, but it opens soon, so if you don't want to miss it, sign up for the waitlist here.​

4) Just respond to this email if you think I can help. I'll either give you my calendar, or connect you with someone who can help.

Dave Talas

P.S: Could you please forward this email to an entrepreneur friend who might benefit from it? They can subscribe here to get the new emails from now on.

You can also read some of my previous emails here.