How I Set Up Claude Cowork the First Time (and Moved My Whole ChatGPT Setup Over)

Most people open Claude Cowork and start typing. I did it differently, and I think that is the whole reason it clicked for me. Before I gave Cowork a single real task, I watched a few YouTube walkthroughs and built a set of workspace folders based on what those videos showed. Following other people’s setups was actually how I learned Cowork in the first place — copy someone’s structure, get it running, then bend it to fit your own work. After that I carried over the personalization I had built up inside ChatGPT. This is the exact setup I landed on, in order, and what I would change if I did it again.

Quick context on why I bothered. I run this blog, I help my wife with her real estate marketing, and I plan a lot of travel. Three different jobs, three different voices, one assistant. Without structure, Cowork forgets who you are between sessions and re-asks the same questions. With structure, it shows up already knowing your rules. That difference is the whole game.

The setup, at a glance	What it does
A workspace folder	One folder Cowork reads every session. This is the home for everything below.
CLAUDE.md (root)	Always-on rules. Auto-loads into every session. Keep it short.
00-About-Me/	Who you are, your rules, your voice, your stack. This is where ChatGPT memory goes.
01-Global-Rules/	Reference files loaded only when their topic comes up.
02-Skills/	Repeatable jobs (blog posts, listings, itineraries) written once, reused forever.
03-Brand-Voice/	Voice rules per brand, plus a list of words to never use.
Projects/	One subfolder per project, each with its own CLAUDE.md.

Step 1: Build the folder before you build anything else

I watched a couple of YouTube walkthroughs first, then came to Cowork with a plan instead of a prompt. The plan was a folder. I made one directory on my machine called Claude-Cowork-Workspace and pointed Cowork at it. Everything else lives inside that folder.

The reason this matters: Cowork automatically reads a file named CLAUDE.md at the root of your workspace at the start of every session. That file is your assistant’s permanent memory. Anything you want it to know without being told again goes in the workspace. Anything you put in a normal chat dies when the chat ends.

My top-level layout, numbered so it sorts in order:

1. CLAUDE.md at the root — the always-on rules.
2. 00-About-Me/ — about-me, my-rules, my-voice, and stack-and-tools.
3. 01-Global-Rules/ — behavior rules that only matter sometimes.
4. 02-Skills/ — repeatable jobs.
5. 03-Brand-Voice/ — one voice file per brand.
6. 04-Templates/ — reusable output shells.
7. Projects/ — one folder per real project.

Step 2: Bring your ChatGPT personalization over

This is the part that paid off most, and it is also where I was wrong at first. I assumed I would have to retype everything from ChatGPT by hand. I was wrong — Claude has a built-in import that does part of the job for you, and I wish I had found it sooner.

The import lives in Settings. Go to Settings, then Privacy, then Memory preferences, and choose Start import. Claude hands you a short request to paste into ChatGPT (it works with Gemini and Copilot too). ChatGPT replies with everything it has saved about you in one block, you bring that back to Claude, and it folds the content into its memory. It takes about a minute and pulls over the useful stuff: your tone preferences, your job and projects, and the context you were tired of repeating. Imported memory can take up to a day to fully settle in, so do not panic if it is not all there immediately.

Here is the catch, and the reason you still build the folder yourself. The import populates Claude’s general memory and your profile preferences. It does not build your workspace — not the per-project CLAUDE.md files, not the brand-voice rules, not the banned-word list, not the Skills. So I treat the import as step zero: run it to get your baseline personalization across, then build the workspace by hand for the project-specific pieces the import will never reach.

What I built by hand, seeded by whatever the import brought over:

1. about-me.md — who I am, my role, my projects, the facts I was tired of repeating. Your ChatGPT answer to “what should ChatGPT know about you” drops straight in here.
2. my-rules.md — how I want Cowork to respond. This is the “how should ChatGPT respond” box, moved over.
3. my-voice.md — how my writing should sound: short and medium sentences, real numbers over adjectives, recommendations stated plainly and then defended.
4. A “never use these words” list. This single trick removes most of the AI-sounding tells. My banned list includes delve, robust, seamless, holistic, moreover, furthermore, in conclusion, and a dozen more. Cowork honors it on every output.

The result is that Cowork writes in my voice on the first try instead of the third. The import gets you most of the way on the personal stuff in about a minute. The folder is what makes it actually sound like you across every brand you write for.

The import brings your memory across in about a minute. The workspace folder is what the import cannot build for you.

Step 3: Add a “Decisions already made” section

This was the best idea I stole from my own frustration. Every assistant loves to re-open settled questions. It suggests the tool you already rejected, the hotel chain you already ruled out, the strategy you already killed. So I added a short section to about-me.md called “Decisions already made.”

It is a plain list of closed questions: which platforms I have standardized on, which I have ruled out and why, the format choices I am not relitigating. Now when I ask for help, Cowork builds on those decisions instead of pitching me alternatives I already considered. It is a small file with an outsized effect on how much time I waste.

Step 4: Turn repeat work into Skills

A Skill in Cowork is a saved instruction set for a job you do over and over. Instead of re-explaining how I want a blog post built every single time, I wrote it down once. I keep six in 02-Skills/:

1. blog-post-writer — long-form SEO posts with proper headings, frontmatter, and image alt text, routed to the right brand voice.
2. seo-meta-generator — title tags, meta descriptions, Open Graph, and schema markup in one block.
3. affiliate-disclosure — FTC-compliant disclosure blocks for any monetized post.
4. real-estate-listing — listing copy in three lengths with Fair Housing guardrails built in.
5. review-responder — short, signed replies to Google reviews.
6. itinerary-builder — trip briefs and full prose itineraries with packing lists.

You do not need six on day one. Build a Skill the second time you catch yourself explaining the same job twice. That is the signal.

Step 5: Give every project its own folder and its own CLAUDE.md

Inside Projects/ I keep one folder per real initiative, and each gets its own CLAUDE.md. The root file holds rules that apply to everything. The project file holds the specifics — the audience, the constraints, the workstreams for that one project. Cowork reads the root rules first, then layers the project rules on top when I am working inside that folder. Clean hierarchy, no contradictions, no bleed between a real estate task and a blog task.

Step 6: The first prompt should be “read everything, then tell me what you understand”

Once the folder was built, I did not ask Cowork to do work. I asked it to prove it understood the setup. My first real prompt told it to read every file in order — CLAUDE.md, then the About-Me folder, then the rules, voice, and each project — and give me back a numbered summary of who I am, what my projects are, my voice rules per brand, the Skills it found, and anything that looked incomplete or contradictory. I told it explicitly: do not write any files yet, just confirm understanding.

That summary did two things. It confirmed the structure was being read correctly, and it surfaced gaps I had not noticed — an unconfirmed setting here, an empty folder there, a rule I had written but never actually pasted into Cowork’s global settings. Fixing those took ten minutes. Skipping that step would have meant discovering the gaps mid-task later.

Step 7: Trim it down so you are not burning tokens

I have to be honest about why I cared so much about this one. In my first week, I burned through my Pro plan’s tokens fast — fast enough that it genuinely worried me — and a lot of my early effort went into making my setup use fewer tokens so I would stop hitting the wall. The cleanup below started as damage control and turned into good practice.

My first version of CLAUDE.md was bloated. I had stuffed every rule into the root file, and a rule that told Cowork to read all four About-Me files at the start of every session. That meant roughly four thousand tokens of context loaded before any session, even ones that did not need my writing rules or my tool stack.

The fix had three parts, and it cut my per-session load by more than half:

1. Split the root file. I moved topic-specific rules out of CLAUDE.md into reference files in 01-Global-Rules/ and 03-Brand-Voice/, and left a small pointer table in the root. Those files now load only when their topic actually comes up.
2. Load on demand, not always. I changed the rule from “read all four About-Me files every session” to “read about-me every session, read the others only when their domain is touched.” A research session no longer pays for my writing rules.
3. Set hard ceilings. I gave each file a line limit — CLAUDE.md capped at 100 lines, about-me at 60 — with a rule that when a file hits the ceiling, the fix is always to compress or archive, never to raise the limit. It keeps the whole system from quietly bloating back up over time.

If you are on the Claude Pro plan, this matters even more than it did for me. Which brings up the install reality nobody mentions.

The install gotchas that cost me time

Three things tripped me up that have nothing to do with the folder structure and everything to do with running Cowork locally.

• Virtualization. On day one, Cowork could not touch my local files. The cause was virtualization being disabled in my motherboard’s BIOS. I told Claude what was happening, it ruled out a couple of other causes, then walked me through enabling virtualization at the BIOS level. The tool fixed its own dependency, which was a good early trust signal — but budget for the possibility that your first attempt does not just work.
• Connector logins are on you. Wiring up the WordPress connector for this site meant I had to log into my host and into WordPress myself, because Cowork would not log in on my behalf. That gate made setup slower and I prefer it that way. Visible friction is the safety feature.
• The Pro plan token wall. On Claude Pro, I was running dry in about fifteen minutes of heavy Cowork work and then waiting roughly five hours for a reset. That is not workable for sustained projects, and it is what drove all the trimming in Step 7. I have since moved to the $100-per-month Max plan, and tokens are no longer the constraint — honestly, they overflow now. The plan is pricey, but the wall is gone. If you are staying on Pro, trimming your context buys you real headroom before you have to make that call.

Build the folder first. Move your personalization over. Make it confirm what it understands. Then start working.

Do this if you are setting up Cowork

• Worth the full setup if you have more than one kind of work, a voice you care about, or any task you repeat weekly. The folder, the About-Me transfer, and two or three Skills will pay for themselves in the first week. This is the configuration that turns Cowork from a clever chatbot into an assistant that already knows your business.
• Skip the heavy version if you are only using Cowork for one-off questions. If you are not coming back to the same projects, a workspace folder is overhead you do not need. Open the chat and go.

If you take one thing from this: do not start with a task. Start with a folder, move your ChatGPT personalization into it, and make Cowork read it all back to you before it does a single piece of real work. Everything good that has come out of my setup traces back to that one hour.

---

Written by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. This is the actual setup running my own Claude Cowork workspace as of June 2026.