Doubao Work adds one-click AI asset migration for conversations, skills, and memory from ChatGPT and Claude Code

Original by AI Woz 2026-10-14 17:25 Zhejiang

It's so good this time that I barely recognize it—again.

There are more and more agents, and everyone is competing hard.

I want to solve an old problem: how do we adapt a new agent to the way we work?

Previously, I would either create a project file containing an agent Markdown document for the agent to read, or prompt ChatGPT to output everything it had recently learned about me and copy that into the new tool's default memory settings.

But neither approach is enough.

Skills, connects—connections to things such as video generation and code management—and the practical experience and work progress you never summarized in project files all get lost during migration.

Even if a model gets a huge upgrade today, migrating manually still takes a long time if its agent has no import option. Otherwise, you can only start by giving it generic tasks such as gathering information.

So when I saw Doubao Work launch one-click AI asset migration, it seemed to bring the losses from switching agents close to a minimum. Specifically, it lets us import our AI assets from ChatGPT and Claude Code into Doubao Work and keep using them. “Assets” sounds abstract; in plain language, it means three things.

Conversations. Skills. Memory.

Conversations are your entire history of talking with AI. Skills are what you have taught it, such as the skills and connectors you configured in Claude or the custom instructions you set in ChatGPT. Memory is the experience the AI accumulates while working with you: it knows your preferred code style, the document formats you dislike, and the pitfalls you have already encountered and should not repeat.

Together, these three things are the accumulated rapport between you and an AI—an asset built through working together.

The process is simple.

Under Data Import and Sync in Doubao Work's desktop settings, it automatically detects which agents are installed on your computer.

On my machine, Claude showed 240 importable items and ChatGPT showed 682.

The import options fall into two categories. Tools and Settings covers skills, connectors, and memory. Projects and Sessions covers conversation history.

I selected both categories and went all in.

It finished in about a minute or two.

Then I tested the result. I opened a past conversation imported from Codex and asked a follow-up directly in Doubao.

It picked up the conversation naturally.

The context carried over. It knew where we had left off and remembered the plan we had agreed on.

But I need to be clear about one limitation.

It can only sync local content: data retained by the desktop client on your PC. If you have always used ChatGPT in a browser and never installed the desktop app, those conversations are stored on OpenAI's servers and Doubao cannot read them.

P.S. You can also choose whether to enable automatic syncing and what it should cover.

Once the move was done, I got curious about some edge cases.

Suppose I tell Claude that I dislike blue-purple interfaces, but tell Codex that I like purple interfaces. If both preferences are imported into Doubao, will they conflict?

I went straight to the Doubao team to ask.

The answer was interesting: it only imports working memory, meaning experience tied to specific conversations and projects. Personal preferences are not imported.

That actually makes sense.

I don't know whether you have tried Codex's feature for importing Claude Code, but it is pretty crude: it simply pulls conversations into Codex. If your Claude Code uses a third-party API, the imported conversations both disrupt the existing sort order and leave the associated file links unusable—even their paths will not open.

Dumping preferences from every platform together indiscriminately would probably create a mess.

So Doubao migrates the most valuable part: experience and context tied to specific work. You can set your preferences again in Doubao Work. It has a separate entry for work-task preference instructions, which also supports importing from other agents.

I also asked about credentials. If Claude Code's AGENTS.md contains global API key and address settings, such as a Supabase database connection string, can Doubao read those too? It does not support that yet, mainly because of concerns about credential conflicts across platforms. I fully agree: keys are not something to cut corners on. With DeepSeek, for example, I used to give API keys arbitrary names whenever I exported them, and later could not tell where the money had gone. At one point, I had GitHub run automatic deployments so often that 100 yuan disappeared in a day or two, with no way to identify the corresponding platform.

Let me also mention a few other features Doubao Work introduced at the same time, all of which I use frequently in Codex.

Continue in a new conversation: branch off from a particular message, carry the context over automatically, and retain the original conversation. If you want to change direction midway through a discussion, you do not need to paste the context again.

Message annotations: select a passage in a reply, add a follow-up comment, and send several annotated passages together. No need to write a long explanation such as “what did you mean by the third point above?”

A conversation-summary sidebar lists the current conversation's tasks, outputs, and skills. After a dozen or so exchanges, you do not have to scroll back to the beginning.

P.S. On the right is a collection of small vibe-coding tools I reverse-engineered using REA. I have handed testing over to Doubao to see whether it can put any application into a comfortable 4:3 viewing layout with one click. That would let even a 13-inch screen show multiple agents' interfaces.

Individually, none of these features is a major update. Together, though, they show Doubao trying to turn AI conversations from disposable interactions into a knowledge asset that accumulates over time.

And finally,

As a programmer, I spent quite a while moving from the terminal to graphical interfaces. I have always thought Codex's interface was good; honestly, it has pushed further on the visual side than its competitors. If Opus 5.5 were not such a good model, Claude's desktop app would already be ten versions behind Doubao. I'm waiting for Doubao Work's models to improve a little more; then I can break out the champagne and perhaps cancel some subscriptions.

After all, integration with Feishu is a huge draw.

Over my month of using Doubao Work, the clearest advantage I have seen is how it presents work for sharing.

With Codex or Claude Code, you might share a GitHub project or a website, but that is not the most convenient approach.

Documents are still what we most often look at when exchanging information, even when the material is meant for an agent.

Feishu documents can hold multimedia in this format and do it better than Google Docs.

Even for video editing, lark cli supports relevant video-understanding capabilities.

Instead of cleaning things up and sharing an entire project for others to understand, we can directly provide a document that most people with ordinary office skills can follow. That lowers the effort needed for others to understand our work progress.

And because Doubao has absorbed my working habits,

I no longer look at a finished document and complain about why several tables were not combined into one,

or why the entire thought process was written into it.

Which brings us back to the familiar routine:

grab a coffee and wait for the agent to finish the work automatically. Nice!

@ Author / Karl

Finally, thank you for reading this far 👏If you enjoyed the article, feel free to support us:Like | Reading | Share | Comment 📣

To receive new posts as soon as they arrive, add us to your starred accounts 🌟