Set Up Claude Like a New Hire

A friend of mine canceled her Sunday afternoon last weekend to "really sit down with Claude."

She's smart. She runs successful teams, advises on businesses, and is good with technology. She blocked the hours, made the coffee, opened the app, and ended the afternoon convinced her Claude was broken.

"It's just being stupid," she told me. "I feel like my Claude is different from everyone else's Claude."

It wasn't broken. She was asking the wrong question.

She'd been asking, "how can you help me?"

That question gets nowhere, every single time.

Asking AI how can you help me is like walking up to a pilot and saying, "where can you take me?" The pilot's honest answer is, "I don't know, anywhere in the world. Where do you want to go?" The question is technically correct, and it also puts none of the work where the work actually belongs.

I've been thinking about this a lot lately, because I keep watching smart people walk away from Claude (and AI tools generally) feeling like the tool failed them. The tool didn't fail them. They handed it a one-way street, and one-way streets always return the same thing: a polite shrug.

Here's the five-part fix I'd give my friend if I were sitting next to her, in the order that actually matters.

1. Tell it about you first

The handoff mental model is wrong. Don't think of AI as a tool. Think of it as a new hire on day one.

You wouldn't hand a new head of operations a wide-open "so what should we work on?" on day one. You'd brief her on the team, the calendar, and the projects in flight. You'd point at the politics and the no-fly zones and tell her where the actual problems are. You'd walk her through how a typical week falls apart and how a good week comes together.

Do the same with Claude.

Walk it through your day. Your morning routine, the tools you use, the team you have or don't have, the work that's frustrating, the work that's flowing, and the things that have been costing you time for months without you noticing.

Then ask, "Based on everything you just learned, where are the top three places you think you can save me time, make me more money, or take work off my plate?"

That question produces a wildly different output. You go from "anywhere in the world" to "here's where we're flying."

2. Pick the right Claude

Claude is actually three products, and which one you pick determines what's possible.

The chat function is what most people are already using. One-off questions, ongoing projects, conversational. The lowest barrier to entry, AIM-days easy. If you've never used Claude before, this is where you start. Don't apologize for staying here.

Claude Cowork is the middle tier. Anthropic built it for non-technical people who want Claude Code's power without the terminal. It can take over your computer, organize files, and run multi-step tasks across your apps. More capability than the chat, less friction than the code option.

Claude Code is the most powerful of the three. It reads your files, controls your browser, and does almost anything you give it access to. Until recently this meant working in a terminal, which scared off most of the people who would benefit most from it. That's changed. Claude Code now runs inside the Claude app. Toggle to code and pick a folder.

The mistake most people make is starting in the wrong tier. If you're doing one-off questions, the chat is right. If you want Claude to act on files, you need at least Cowork. If you're trying to systematize how an entire business runs, you need Code. Pick the surface that matches what you're actually trying to do.

3. Stop typing

The slowest way to use AI is by typing.

Open Claude and hit the microphone button. If you're working in an IDE or want more control, use a system-wide dictation tool. I use Wispr Flow and Monologue depending on the context.

Then talk.

What to talk about on day one: yourself.

Here's the exact prompt I give Claude in voice. "I'm going to tell you about who I am, what I do, and how I like to work. At the end, ask me up to 20 follow-up questions in a batch to fill in the gaps."

Then I just go for 15 to 20 minutes, hitting why I went into marketing, why I moved to New York, how my team operates, what kind of content I make, the work I hate doing, and the work I'd happily delegate tomorrow if someone capable showed up. (Same brief I'd give a new head of operations, by the way. The point is the brief. The format is interchangeable.)

When Claude sends back its follow-up questions, I answer those in voice too. By the end I have roughly 30 minutes of monologue compressed into a project. Then I tell Claude to turn it into a permanent set of project instructions: "always create content like me," "always plan my week the way I would," "always reply to email in my voice."

At this point Claude has graduated from chatbot to a colleague who knows me.

Two reasons this works. Voice is roughly 3x faster than typing, which means Claude builds a richer model of you before you ever ask it for anything useful. And voice catches your actual cadence. Written-you and spoken-you are different people. Spoken-you uses fillers, side jokes, asides, the way you actually think out loud. Claude needs the spoken version, because the version of "you" that Claude eventually produces will sound like whichever version of you it learned from.

The output from a Claude that's heard 30 minutes of you is a different planet from one that's read 30 minutes of you.

4. Plug it into the rest of your life

This is the moment Claude stops being a chatbot and starts being an operator.

The bridge between Claude and the rest of your stack is called MCP, which stands for Model Context Protocol. Anthropic built it so Claude (and other LLMs) can interface directly with the tools you already use: email, calendar, project management, CRM, design files, and almost anything else you can think of.

Open Claude, go into Connectors, and add the tools where your work actually lives. Mine are Gmail, Google Calendar, Google Drive, ClickUp, Figma, and Slack. If you don't use those tools, the principle holds. Connect the ones you do.

When you grant access, Claude asks for read or write permission per tool. Keep most read-only. For the ones where you want drafting (email, Slack), give write. Never auto-send. The point of having an assistant is that you stay in the loop on what goes out.

There are three things I run every single week, and all three used to cost me real hours.

The first is email triage. Every morning Claude runs this prompt: "Go through my inbox. Flag what needs my attention. Archive the junk. Draft replies to anything I owe a response on, but don't send." I review the drafts, edit the ones that need it, and send the rest. Roughly an hour back per day.

The second is a cross-tool weekly digest: "Summarize my Slack, email, and ClickUp from the last week. What's the team working on? What's blocked on me? What's slipping?" This is the read I wish I could write for every founder I know. Visibility without the meeting.

The third is Sunday planning: "Go through my calendar for the week, my outstanding ClickUp tasks, my unread Slack, and my flagged email. Tell me what should be on my plan, where to put focus time, and which meetings I should cancel or hand off." I review, decide, and ship the week.

One rule, written on the inside of my eyelids: garbage in, garbage out. Connect the tools where the real work lives, not the ones that look impressive. Claude is a context machine. Feed it good context.

5. Build the structure that scales

I run a deliberately lean agency with a full client roster. The reason we can do roughly five times the work we used to with the same small team comes from AI plus the right context architecture, working together.

Here's the architecture.

One top-level folder per client. Inside, structured subfolders for meeting notes, intake documentation, planning docs, sales page copy, email copy, social posts, and survey data. Everything from the working relationship lives in a known place.

At the top level of every client folder there's a README. The README does two things: it explains how we work with that client (their preferences, their no-fly zones, their internal vocabulary), and it points Claude into the right subfolder for the task at hand.

The folder syncs to GitHub daily. It could also just live locally. The key is the structure, not where it lives.

People sometimes ask why I don't use Claude Projects for this. I tried. With all the info dumped in, Claude wasn't reliably reading every file. With structured folders and a README pointing the way, Claude knows where to focus, and the signal-to-noise stays clean.

Once the structure exists, the asks get interesting. "Review this product idea. Make sure it fits within Client X's product plan." Or, "Client wants to change direction on this launch. Read the last six months of meeting notes and the launch brief, then propose a revised plan." Or, "Surface every story from this client's meeting notes that touches their founder journey."

The work is organized, and the organization is what makes the work possible. The folder structure is the moat. The model is the engine.

We still put a human touch on everything that ships. The bench of context Claude is working from, though, is a thing no human could hold in their head. That's where the actual scale comes from.

---

So that's the five-part fix.

The version of Claude my friend met on her Sunday afternoon (the one that "felt different from everyone else's Claude") was a Claude that had been given no context, no architecture, no voice training, no connectors, and no folder system to read from. Of course it felt broken.

The version of Claude I use to run my company is the same product, with different context, a different architecture, a different surface area, and the same model behind it. Context is the moat.

This is good news, actually. The version of Claude that works for me isn't a function of being early or being technical or being in the AI-bro internet. It's a function of how I set it up, which is something anyone can do in a weekend.

The first weekend is the one nobody tells you about. Spend that weekend on context. The output comes the second week.