How I got my files out of Claude Cowork and into Claude Code

TL;DR: The Claude Code test drive is a free paste-in prompt that reads a folder of your files, interviews you, and tells you what moving to Claude Code would actually take, as of August 2026. It’s read-only and it’s not gated: no email, it’s just there. The finding that started this page, from my own account as of August 2026: the files I uploaded to Claude have a download button, and the files Claude wrote for me don’t. My data export turned out to be the mirror image: it returned Claude’s documents from my chat Projects and skipped my uploads. I’m not here to talk you into switching, and there’s nothing to buy at the end.

Watched the video? Jump straight to the test.

Why I moved

I moved because I kept hitting walls in Claude Cowork, and that story is already on this site: the full comparison, including who should stay, is its own post. This page is about the part the comparison didn’t cover: the move itself.

When I decided to start using Claude Code, my work was stuck all over the place. Threads in Claude chat. Files in chat Projects. More files in Cowork projects. Some things on my computer, the rest in my account’s cloud.

Then my Cowork video went up, and the comments kept asking one question: “how did you swap everything over?” So this page is that answer: where your files live, how I got mine out, and a test to run on yours.

Where your Claude files actually live

Some of your work sits in folders on your own computer, some sits in your Claude account’s cloud (chat Projects and Cowork projects), and some is buried in the conversations themselves. The part on your computer, including anything Cowork’s local mode or connected folders already put on your disk, is ready; it doesn’t have to move anywhere. The cloud is where I found the catch.

I went looking in my own account in August 2026, in some old projects. The files I had uploaded, I could download again. One at a time, but the button is there. The files Claude wrote for me in those projects? I could open them. I could delete them. I could not find a download button. Selecting files gave me one control: a trash can. And I never found a download-everything button either.

Your uploads probably still exist somewhere on your disk. The files that live only in the cloud are exactly the ones you and Claude made together: the plans, the drafts, the strategy documents. Those are the ones that need rescuing, and they’re the ones without a button.

Get your files where the test can see them

There are four ways out, and I’ve tried all four myself, as of August 2026. Your uploads are the easy case: the file list downloads them. Everything Claude wrote needs one of these.

1. Ask a session to hand your files over. Open a session in the old Cowork project and say: give me my files. It reads each one and writes you a copy, and it can bundle everything into one zip. Know what’s happening under the hood: it rebuilds every file from scratch, retyping every character, which costs usage and fills up its memory. Do big projects in batches, a handful of files at a time, and spot-check the copies against the originals. I did this on eighteen files, and the ones I checked looked okay.
2. Have it write straight into a folder on your computer. This is Cowork’s trick: if you’ve connected a folder from the desktop app, tell the same session to put the files in there. I tested this on my own machine, and the file landed exactly where it was supposed to.
3. Copy everything out by hand. Open each file, copy, paste into a document on your disk. It works everywhere, it costs nothing, and depending on how much you have, it’s a lot of work.
4. Request your data export. In your account settings, under Privacy, you can ask for your data. Anthropic emails you a download link that’s good for a day. Request it on a computer, not in the phone app, and know that a big export can arrive as more than one zip; mine was labeled batch zero.

The export is how I rescued my chat Projects, which have the same one-way door as Cowork: uploads return, Claude’s documents don’t. Once you have the zip, point Claude Code at it and ask it to pull out the files Claude wrote for you in those projects. For me it found them all; a quick check against my project list confirmed it, and I’d do the same check on yours. It pulls the files’ content out of the sessions and rewrites them onto your disk. Still token-intensive, but Code gives you a better picture of context and token use than the chat tools, so you can monitor it as you go. And you don’t have to open every project separately to pull files out one by one. Keep your own originals of anything you uploaded; the export won’t give those back. None of this is documented anywhere official that I can find, so if your export behaves differently, tell me in the comments on YouTube.

Some of what you’ll want to keep never became a file at all. It lives in old conversations: decisions, plans, things you worked out in a chat and never saved. Getting those out has its own catch. In a long chat, scrolling back and copying by hand won’t reach everything. You can ask the old session to write you a handoff file instead, everything important in one document, but read it before you trust it: a long conversation outruns its own memory, and the session doesn’t always know what it can’t see anymore, so it writes you a confident summary of the part it remembers. The complete record is the data export. The archive holds your whole chat history, including what you can’t scroll back to, and a fresh session can dig through it, searching for what matters instead of reading every line, and pull out what’s worth keeping. One caution before you hand it over: the archive is your entire history. Only put it in front of the test, or any agent, if you’re comfortable with it reading all of it. If you’re not, have a separate session dig through it first and bring over only what it finds. I also did this dig with my old ChatGPT history, to pull out things I knew were in there that I wanted in my knowledge base.

The Claude Code test drive

Version 1.0, August 2026. The test is free and it’s plain text: you can read every word below before you run it.

What would Claude Code do with your files?

You don’t have to switch to find out. This test answers one question: if you pointed Claude Code at your files today, what would it find, and what would moving actually take?

Nothing gets changed. The test reads, asks you a few questions, and gives you a written verdict. Whether you switch afterward is still entirely up to you.

The test changes nothing for a reason bigger than caution. The moment you start working in two places at once, you have two versions of the truth, and they drift apart. So: test first, decide once, move once.

If you’ve migrated platforms before, and it was a lot of work, and you didn’t really know what you were getting into, this is for you. I moved my own work out of Claude Cowork projects and Claude chat Projects, and the move was messier than it needed to be because I didn’t really understand at the time where my files were. When I was building my knowledge base I ended up with a lot of duplicate and divergent files, and it wasn’t clear what the canonical versions were. This test is the look-before-you-leap I wish I’d had at the time.

You also don’t need to arrive knowing what you want from Claude Code. The test carries a menu of things people run, pulled from my own working system, and asks which of them you’d want. Not knowing yet is a fine answer.

Before you start: Claude Code comes with the paid Claude plans (Pro and up). If you’re on the free plan, that’s the real cost of taking the test.

Step 1: put everything in one folder

Gather the files you’d want Claude Code to see into one folder on your computer. Don’t go out of your way to sort them. Sorting is Claude’s job, and the mess is part of the test. (If they’re already sorted, leave them sorted.)

If some of your work is still stuck in the cloud, in Cowork or in chat Projects, get copies onto your own disk first. The page this test lives on walks you through every route out, the same ones I used for my own move. Mostly just chatted? That counts: the page also shows how to turn conversations into files you can bring. And if your folder ends up almost empty, take the test anyway. It measures where you’re starting from, and “nearly nothing” is an honest place to start.

Step 2: take the test

Open Claude Code in that folder and paste this:

I'm thinking about moving my work into Claude Code, and this is only a
test run. You are read-only today: don't create, edit, move,
or delete anything. If something seems worth fixing, put it in your report
instead of fixing it.

Start with one calibration question and wait for my answer: how have I
used Claude so far? Chat only? Chat with uploaded files? Projects? Cowork
with connected folders? Something else? Scale everything that follows to
my answer. If I'm new, use plain, everyday words and explain any term a
newcomer wouldn't know. If I clearly know my way around, skip the
explanations.

Then look around and make an inventory: what kinds of files are here, how
they're organized, what belongs together, and where you see more than one
version of the same thing. Flag anything that looks outdated. Old files
matter to me: a stale file that still gets read can quietly steer things
wrong later. If the folder is nearly empty, say so without judgment: a
small pile is a fine starting position, and the test works the same. If
there's too much here to read everything, don't pretend: sample the
newest files and the ones other files mention most, and tell me plainly
what you didn't read.

Then interview me, one question at a time. Wait for my answer before
asking the next one.

1. What work do I do in a normal week, and which parts of it happen at
   this computer?
2. What does my current AI routine look like: which tool, how often,
   for what kinds of tasks, and what would I miss most if it disappeared
   tomorrow?
3. What do I wish my AI already knew about me and my work, without me
   explaining it every time?
4. What is annoying or broken about how I work with AI today?
5. Which files or folders are precious: never edited, never reorganized,
   never touched without asking me first?
6. When new material arrives (downloads, notes, attachments), where does
   it land, and does it ever get sorted?
7. Here are things people run in Claude Code. Which of these would make
   me sit up, and what's missing from this list for my work?
   - a memory that stays current: notes about me and my work that it
     updates as we go
   - a work log it writes: what happened, what was decided, what's next
   - a friction list: I name what annoys me, it proposes fixes
   - an inbox routine: I drop files into one folder, it sorts and files
     them
   - a voice-memo pipeline: I talk, it transcribes, files, and acts
   - a dashboard it rebuilds, so I see my projects at a glance
   - harvesting: it mines my finished work for reusable pieces
   - house rules: my corrections become standing rules it follows in
     every future session

One rule for our whole conversation: if I ask whether something is
possible in Claude Code, Claude Cowork, or any other tool, don't answer
from memory. This space changes monthly and your training may be out of
date. Check the current documentation with your web tools before you
answer, and if you can't check, say plainly that you can't verify it
today.

Then give me your verdict in five short parts, in plain language:

1. What would work in Claude Code exactly as it is today.
2. What you'd want reshaped before it earns its keep, one sentence of
   "why" each.
3. The duplicates and probably-stale files you found, and which version
   looks like the real one.
4. A starter structure: the folders you'd suggest, one line each on what
   goes where. And if my files are a pile you'd have to re-read
   wholesale every session, say so, and describe the alternative: a
   short index file the AI reads first, plus linked notes it follows
   only when a task needs them. Suggest that only if it fits what I
   said I want.
5. My first three sessions: three small, concrete tasks, easiest first,
   that would make the move real. Pick at least one from whatever made
   me sit up in question 7.

Don't flatter me. If this pile isn't ready, say so, and say what "ready"
would take.

One exception to read-only, at the very end: ask me whether I want this
verdict saved. If I say yes, write it to a single new file named
migration-test-verdict.md in this folder, so my next session can pick
up where we left off. Create nothing else, change nothing else.

Answer its questions honestly. The verdict is only as good as your answers, and nobody’s grading you.

My recommendation: dictate your answers. Typing is slow, we all stop to fix our mistakes, and most of us can’t get a thought out through a keyboard anywhere near as fast as we can say it. One caution: don’t dictate straight into the chat box unless your dictation tool saves your answer as you go. It is so frustrating to get deep into an answer and lose it all because the Wi-Fi glitched or the session timed out. My version of Word has a microphone button; I dictate there and paste it over.

What to expect while it runs: Claude Code reads freely in the folder you opened it in, and asks permission before anything beyond that. During the test it may ask to search the web (that’s the checking-the-docs rule doing its job) or to run a small command to open a file type it can’t read directly, like a Word document. Those are fine to allow. It should never ask to change, move, or delete a file today. The single exception comes at the very end, and it’s the good kind: the test offers to save its verdict as one new file in your folder. If it asks to change anything else, say no and point it back to the first line of the prompt.

You also don’t need to change any settings or pick a special model for this. Whatever your plan runs is fine; the test only reads and asks.

Step 3: read your verdict like a mover’s estimate

Two warnings from my own move.

You will find duplicates. I arrived with multiple versions of half my important documents and no idea how they overlapped. That’s normal. Finding the overlaps is exactly the kind of cross-reading Claude Code is good at and you are slow at. On moving day, make “pick one canonical version of each” its first real task. That goes double if you go on to build a system that runs on these files, whether it’s something like mine or an adaptation of what you already have: duplicates left standing get baked into whatever you automate.

Old versions don’t retire on their own. Don’t assume that a note saying “this version is outdated” will actually stick. Your AI might open an old file tomorrow with no idea the contents are dead, and stale content can drive real actions. When you migrate for real, have Claude delete the old versions or move them into an archive folder your setup doesn’t read. I learned this the slow way.

Nervous isn’t a stop sign here, it’s a checklist: keep client files and anything private out of the test folder if you’re not ready, make sure everything is backed up first, and hand the test copies instead of originals if that helps you sleep. In security language, keep your blast radius small: set things up so the worst thing that could go wrong is small enough that you’d shrug. I made a whole video about doing exactly that, and its message fits in one sentence: you can try all of this without risking anything you actually care about.

One thing you should know: the test can tell you what moving might look like, but it can’t tell you whether you’ll actually like working this way. That part still takes a first week, and the test’s “first three sessions” list is designed to get you started.

And go in knowing the trade-off. A system like this takes maintenance to keep its content fresh, and one day you may catch yourself running a system you don’t fully understand, answering questions you don’t feel qualified to answer. That’s the strange part of this era: we get to run things we couldn’t have built ourselves. It lets you do more than you otherwise could, and it costs some of that feeling of full control.

In the end, everything is just files. That’s the whole secret, and it’s also why you’re able to test the water before you dive in.

Is the test safe? I tried to break it

The test is read-only, you can run it on copies, and it survived two sandbox runs designed to trip it. One thing to know: what the test reads ends up in your chat history on the server, like any Claude conversation. I was hesitant to write up a prompt and tell everybody: go run this. So I built a sandbox: invented users, each with a folder seeded with real-life mess. Duplicate files. To-do lists that contradicted each other. A tools list that was out of date in a quiet, hidden way. Then the test ran on those folders, twice.

It caught every trap I planted, in both runs. It kept its hands off the files, declining even to write its own memory. And when it hit a question about what these tools can do today, it didn’t answer from memory: it went and checked the current documentation. On your machine, that’s the moment it asks your permission to search the web. When I reran the final version, it stayed read-only through the whole interview and wrote a single file, the saved verdict, only after being told yes.

Here’s the part I find genuinely cool. In one of those runs, the verdict came back: for this user, Cowork is probably the better fit. Good. A test that can only say yes is a funnel. This one can come back negative, which is the whole reason it’s worth taking.

Of course, invented users aren’t real people, and this was my own little at-home sandbox, not a lab. The traps were caught every time; the extra observations beyond them changed between the two runs, so don’t count on those. Let me know how it works on your real files. I’m curious.

Can you go back?

Yes, at any point, because the whole move is copies. Nothing gets deleted: your Cowork projects and your chats stay right where they are until you decide otherwise. You can take the test today, read your verdict, and keep working in Cowork tomorrow.

Is Claude Code only for programmers?

No. I’m 51, I’m a freelance science writer, and I don’t code; I talk to Claude Code in plain English every day (and sometimes in really bad German), and I get more from it than from any other AI tool I’ve tested. People keep telling me in the comments that Claude Code is for programmers and that I’d be just fine using a different system for my writing. As of this August, even Google’s AI Overview calls Claude Code a tool “for developers.” But here’s the thing: you don’t have to know how to code to use it. You have to know what you want, and how to ask for it in plain words. What learning this at my age actually looked like is its own story.

What to do with your verdict

Take the verdict seriously in both directions.

If it says the move would pay, start with the three tasks it wrote you. When you’re curious where this can go, my homepage has the map of what I run today. The test also lives on the downloads page, next to the other files I hand out, and it will still be free when your files are ready.

And if it says stay in Cowork? Believe it. A lot of people are exactly where they should be, and nothing about your setup has to change. Then you found out for free.