Migrating AI Context to awork

Here's how to transfer instructions, knowledge, skills, and saved information effectively from your LLM to awork - without bringing along your entire chat history.

Key benefits and use cases

You don't need to transfer your chat histories from an LLM to awork. In fact, that would be counterproductive: chat histories contain valuable information alongside rejected ideas, corrections, misunderstandings, outdated numbers, and small talk. A new agent can't reliably distinguish between these, so it might fall back on information that's long since been superseded.

What's actually worth transferring is much more compact:

• Project instructions and personas

• Knowledge files and other already structured content

• Skills and recurring workflows

• Saved information about how you work

• Recurring rules and preferences

You can take much of this over directly. A full export often isn't necessary at all.

How does the transfer to awork work?

Here's how to map your existing content to the right destination in awork:

The Workspace Context is particularly important: it's maintained once for the entire workspace and is available to all agents. Company-wide knowledge belongs there instead of being copied into individual agents. You need workspace settings permissions to edit it. Designate a responsible user for this.

The quick version in 30-60 minutes - without export

This gets you roughly 80% of the results - with no export needed. Only do a complete export if something's still missing afterward.

Step 1: Import existing skills directly

If you've written Claude skills, you can import a single SKILL.md file or a skill package directly. No need to rewrite them.

Step 2: Take over already structured content

Open your project or custom assistant and transfer:

1. Project instructions / System Prompt → Insert into the awork agent's System Prompt

2. Knowledge files → Attach as files to the agent or include in the Workspace Context for company-wide content

You've already created and reviewed this content. It typically doesn't need additional editing.

Step 3: Transfer memory

Memory, your LLM's recollection, isn't part of the export. But the assistant can read it out. So start a new chat in the same account and use this prompt:

List everything you have stored in memory about me and my work.

Include: my role and background, my company and products, my recurring
responsibilities, my preferences for how you respond, explicit corrections
I have given you, terminology I use, and any people or projects you know about.

Mark each item with where it came from:
- [memory] something you have actually stored about me
- [chat] something you retrieved from an earlier conversation
- [inferred] something you are concluding rather than reading

Rules:
- Only list what you can actually access right now. If memory is unavailable
  or empty, say so plainly instead of reconstructing it from this conversation.
- Never claim to have read memories or past chats you cannot actually access.
- Do not turn a question I once asked into a lasting preference.
- Do not include passwords, API keys or access tokens.

At the end, state briefly what you were able to access and what you could not.

Output as a plain Markdown list, grouped by topic.

Check the results afterward and remove any information that's incorrect or outdated. Split the remaining content into:

• Personal work preferences → transfer to your own agent

• Company facts → include as Workspace Context

Step 4: Create a document for recurring tasks

You don't have to work through your entire chat history. Focus on the five to ten tasks you regularly use an assistant for.

Open the relevant project or thread and ask the assistant this for each recurring task:

I am moving this work into a new AI assistant that has access to our
project management system.

Produce a single handover document for the recurring task "<name the task>",
using exactly these sections:

## Background that always applies
## How I want this done (rules, format, tone, banned wording)
## Decisions already made — each with the date it was decided
## Still open / undecided
## No longer valid — things we tried, changed our mind about, or dropped

Rules:
- Only include what is actually supported by this project or conversation.
- Do not turn your own past suggestions into decisions I accepted.
- Do not turn a question I once asked into a lasting preference.
- Date anything that changes over time: numbers, prices, deadlines, statuses.
- If something is contradictory or you are unsure, put it under "Still open"
  rather than guessing.
- Leave out anything that is simply current project status, task lists,
  client lists or who is working on what. The new system has that live.

Step 5: Review content before uploading

Check every document before uploading. Pay special attention to:

• Decisions: Was this actually decided?

• Currency: Is this information still actually current?

• Preferences: Does it apply generally or only to a specific task?

• Company data: Should this information leave your organization?

Correct the information directly in the document.

Step 6: Remove what awork already knows live

Many old context pieces are just static copies of data that awork already manages currently. This includes things like client lists, project status, deadlines, and responsibilities.

Remove such content from your handover documents. Let the agent use the live data from the system instead. This prevents a second, outdated source from being created and delivers real added value.

Step 7: Put each piece of content in the right place

Use the mapping from the table above. As a quick guideline:

• "When X happens, do Y" → Skill

• Information for only one agent → File on the agent

• Information for the entire company → Workspace Context

If multiple people are migrating, not everyone should maintain their own version of company facts. Designate one person responsible for the Workspace Context.

Skills can be shared across the workspace or kept private. Make that decision consciously for each skill.

Tip: Give every document a date and a responsible person.

Step 8: Verify with real work

Take five real tasks from the past few months where you still have good results from your LLM. Run them with the awork agent again and compare the results.

Use the results to determine what additional information the awork agent should be given instead of uploading more material as a precaution.

Then plan for two to four weeks of small fixes. Keep a list of adjustments you need to make repeatedly, and regularly incorporate consistent rules into an agent's System Prompt or the Workspace Context.

Alternative: Use the complete export from ChatGPT or Claude

Export from ChatGPT or Claude

1. On Claude: open Settings → Privacy → Export data or on ChatGPT: open Settings → Data Controls → Export Data
   Both providers send you a download link by email. (Note: the link expires.)

2. For team or enterprise plans, contact your administrator since individual members typically can't start an export.

3. Unpack the download. You'll find one or more large JSON files with all conversations. These are unsorted. Some exports include a manifest file that only contains download links. Use the actual conversation data.

4. Find the conversations you need and process them individually with the handover template from step 4. (Don't just summarize the entire export as one thing.)

5. Keep the complete export as an archive outside your systems and set a deletion date.

Important notes

So make sure:

• The Workspace Context remains a shared source

• Personal preferences stay separate

• Skills are consciously shared or kept private

• Documents have an owner and a date

• Results from the new awork agent are verified against real tasks

• Recurring fixes are regularly converted into consistent rules

Handling company data

An export can contain customer names, contract details, internal decisions, and personal data of colleagues and third parties.

Transferring it to another system is therefore a data transfer. On a company account, that's not a decision individual employees should make alone.

FAQ

Can I import my ChatGPT or Claude conversations directly to awork?

No. Conversations can't be imported directly. Taking over raw data from the chats wouldn't make sense anyway, since they also contain rejected, outdated, or confusing content.

Can I export just specific conversations?

No. Providers typically export your entire account history as a bundle of JSON files. Individual conversations can't be selected during export.

Where does company knowledge belong?

Company-wide facts belong in the Workspace Context. Content that only one agent needs goes as files attached to that agent.

How long does the migration take?

The quick version takes about 30 to 60 minutes according to the guide and delivers roughly 80% of the results. Afterward, you can typically expect two to four weeks of small fixes.

What should I do with the complete export?

Only use it if you need information from old conversations that doesn't exist anywhere else. Process the conversations you need deliberately and then keep the raw export as an archive outside your systems.