By Robiul Islam — Founder, ExpertsCircle.io
Why I Switched
I used ChatGPT heavily for about two years. Product decisions, architecture, legal work, financials. And it was fine — until "fine" became the problem.
The responses felt confident but hollow. It agreed with flawed assumptions instead of catching them. I'd ask it to critique my approach and it would validate the core idea even when the core idea was wrong. Claude pushes back. When my thinking has a gap, it says so. That difference compounds over hundreds of sessions.
Then there's what Claude can actually do. Cowork mode has no equivalent in ChatGPT — it accesses your local files, runs scripts, controls a browser, and chains multi-step tasks autonomously. On top of that, Skills let you extend what Claude can handle in a session: legal review, Excel, PowerPoint, scheduling — all triggering automatically without switching tools or context.
On Projects: yes, ChatGPT has those too. But Claude's lets you load knowledge files and write detailed instructions that shape every conversation in that workspace. Context that actually sticks, from message one.
Better reasoning. A tool that can act, not just answer. Skills that extend it further. That combination was enough. The only problem was two years of conversations sitting in ChatGPT that I didn't want to lose.
The Migration (Step by Step)
I did the whole thing through Cowork. It took a few hours the first time because I was working it out as I went. With these prompts, you could probably finish in an afternoon.
What you get at the end
Each Claude Project has a knowledge file (your ChatGPT history, summarised), project instructions (so Claude knows your context from message one), and anything that didn't fit a project stays searchable in a local archive.
Step 1: Export from ChatGPT
ChatGPT → Settings → Data Controls → Export Data. You'll get a ZIP by email. Save it somewhere easy to find.
One thing to know upfront: ChatGPT does NOT include your project instructions in the export. I assumed it would. It doesn't. The system prompts you carefully wrote? Gone. Just empty strings in the JSON. So you'll be writing those fresh for Claude. (Honestly, mine turned out better the second time around.)
Step 2: Let Cowork process the export
Open Claude desktop, start a Cowork session, select the folder with your export ZIP. Then:
"I have a ChatGPT export ZIP file in this folder. Please unzip it, parse all the conversation JSON files, group the conversations by topic or project area, and create a separate knowledge summary markdown file for each group. Each summary should list the conversations with their titles, dates, and a brief description of what was discussed."
Claude will chew through the files, figure out what goes where, and produce a summary doc per project area. Mine had 3,674 conversations across 37 JSON files and it sorted through all of them.
Step 3: Check the groupings
Claude will come back with something like: "I found 5 project areas..." and list them out. This is your chance to fix anything.
"That looks right, but split [Project B] into two — one for [topic X] and one for [topic Y]. Everything else goes into a General Archive folder."
Don't force everything into a project. Miscellaneous stuff, random one-off questions, experiments — let those go to the archive. You can still search them later.
Step 4: Create the Claude Projects and upload
Go to claude.ai → Projects → New Project for each group. Then back in Cowork:
"Please upload each knowledge summary file to the matching Claude project. The projects are already created on claude.ai."
Cowork opens the browser, navigates to each project, and uploads the files. You'll see them land in the Files section of each project.
Step 5: Write the project instructions
This is the step that matters most. The knowledge file gives Claude your history. The instructions tell it how to actually work with you. Without instructions, you're still going to be explaining yourself at the start of every chat.
"Based on the knowledge summaries you just created, write project instructions for each Claude project. Cover: who I am, what the project is about, what terminology or formats I use, the level of detail I expect, and how Claude should approach responses."
Read the drafts. They'll be 80% right. Fix the rest:
"For [Project Name], add that I want direct answers, no preamble. Remove the part about [X]."
Then:
"Save these instructions to each Claude project now."
Cowork opens each project, drops the instructions in, and saves them. Done.
Step 6: Make the archive searchable
For everything that didn't make it into a project:
"Create a Cowork task for searching my ChatGPT archive. It should search all the archived conversation folders by keyword and return matching conversations with context."
Now whenever you need something from your old history:
"Search my ChatGPT archive for conversations about [topic]"
Nothing is lost. It's just filed.
Was It Worth It?
Yes. What I have now is better than what I was running in ChatGPT.
Every new chat starts with context already loaded. I don't waste the first few messages getting Claude up to speed. The project separation means each workspace has its own tone and focus. And the instructions I wrote from scratch are tighter than the ones I'd been patching together in ChatGPT for months.
The one thing I'd tell you before starting: budget proper time for Step 5. Writing good instructions is where the real value comes from. Everything else is just file management.
Robiul Islam is the founder of ExpertsCircle.io — a recruitment marketplace that connects hiring managers directly with experts, cutting out agencies entirely. Follow him on LinkedIn.