ChatGPT → Claude Migration

You are guiding a user through a complete migration from ChatGPT (or any AI assistant) to Claude. The goal: by the end of this process, Claude will know who they are, how to communicate with them, and what they're working on — without needing to re-explain in every conversation.

This is a structured but conversational process. Read the full skill before starting.

What you're building

You'll produce two things:

1. A set of context documents — markdown files the user loads into Claude Projects
2. A folder structure — organised so the user can find and update their context over time

The context documents are:

• GLOBAL_CONTEXT.md — identity layer, communication rules, beliefs (loaded by all projects)
• One [ROLE]_CONTEXT.md per role the user has identified

Phase 1: ChatGPT Export (skip if user has no ChatGPT account)

Before the interview, walk the user through exporting their ChatGPT data. This gives you raw material to work from and means nothing gets lost.

Read references/export-guide.md for the step-by-step instructions to give the user.

Key points to communicate:

• The export takes up to 24 hours to arrive by email — they should start it now
• You don't need to wait for it to proceed — move straight to Phase 2
• The most valuable file is conversations.json — it contains everything ChatGPT knew about them

Where to save it: Tell the user to save the export to the folder they've selected in Cowork (their workspace folder). This is important — Claude can only read conversations.json if it's in the workspace folder. If it's saved elsewhere, Claude won't be able to access it for cross-checking.

When Claude reads conversations.json: Only use it as a cross-check after context docs have been generated — not as the primary source. conversations.json can be very large (100MB+); do not attempt to read it in full. Instead, scan selectively for:

• Recurring topics, domains, or projects that didn't surface in the extraction output
• Vocabulary and phrasing the user uses for their own concepts
• Named people who weren't captured in Section 9
• Decisions made that didn't appear in roles or current state
• Any role or use case the extraction prompt may have missed

If the user already has the export saved in their workspace, ask them to confirm the filename and location so you can access it when needed.

If they're migrating from a different AI tool (Gemini, Copilot, etc.), skip the export step and proceed directly to Phase 2.

Importing memories from their previous AI: Claude has a dedicated memory import tool at claude.com/import-memory (also reachable via Settings → Capabilities → Memory → Start import). It generates a prompt the user pastes into their old AI, exports everything that AI has stored about them, and they paste the output back into Claude to create individual memory entries.

If the user wants to import memories:

• Direct them to claude.com/import-memory — the tool walks them through it
• Note that it's marked experimental and may not retain all items; it prioritises work-related

content over personal details

• Memory import is fast and useful, but it captures shallow facts — not a substitute for the

full extraction prompt this skill uses. Recommend doing both: memory import for a quick baseline, this skill for the deeper context that actually changes how Claude works with them

Folder to recommend: Suggest they create a folder called Claude Migration (or AI Migration) inside their selected Cowork workspace folder — this keeps the export, context docs, and migration materials in one place that Claude can always access.

Phase 2: Context Extraction via AI Self-Interview

Rather than asking the user a series of questions directly, you'll give them a structured prompt to run inside ChatGPT (or whichever AI they're migrating from). That AI already knows them — let it do the work of surfacing the context. The output comes back to Claude, which then builds the context documents.

This keeps the process fast and draws on existing relationship history rather than starting cold.

2.1 — Give the user the extraction prompt

Tell the user (adapting the AI name to whichever tool they're migrating from — ChatGPT, Gemini, Copilot, etc.):

"We're going to get [their AI] to summarise everything it knows about you — rather than me asking you a bunch of questions. Copy the prompt below and paste it into [their AI]. Then paste the output back here and I'll build your context files from it."

Provide this prompt for them to copy:

I'm migrating to a new AI assistant and need you to help me carry my context across. Please
write a comprehensive self-description covering everything you know about how I work, who I am,
and how I like to be spoken to. Structure your response with these sections:

1. WHO I AM
   - My name, location, and what I actually do (not just job title — what I'm trying to achieve)
   - My language preference (UK/US English, or other)
   - Relevant professional background that shapes how I think
   - What my MBTI and DISC profile is (what you have been told, as well as what you can infer from past interactions with me)
   - My desires
   - My aspirations
   - My fears
   - My frustrations
   - What you can describe about me from all our past conversations, especially things which I may not know about myself
   - What I would say to my younger self, based on our past conversations, what you know about me and what you can infer about me.

2. HOW I LIKE TO BE SPOKEN TO
   General preferences:
   - My preferred tone (formal/casual, warm/direct, collaborative/instructional)
   - My preferred response length and density (brief and punchy vs. thorough and detailed)
   - How I feel about structure: headers, bullet points, numbered lists — do I like them or find
     them excessive?
   - What a great response sounds like to me — use specific examples from our history if possible

   In different circumstances:
   - When I'm brainstorming or thinking out loud — what do I want the AI to do? (Push back?
     Build on it? Ask questions? Just follow the thread?)
   - When I'm making a decision — do I want options laid out, a recommendation, or just analysis?
   - When I'm under pressure or moving fast — short and direct, or still thorough?
   - When I'm asking for feedback on my work — how honest, how blunt, how much encouragement?
   - When I'm learning something new — do I want it explained simply or treated as capable of
     handling complexity?

   In my own words:
   - Quote anything I've said directly about how I want to be spoken to — e.g. "stop
     over-explaining", "just give me the answer", "don't hedge so much", "treat me like a peer"
   - Note any frustrations I've expressed about AI communication style, even in passing

   Anti-patterns — things I dislike:
   - Response patterns that have annoyed or frustrated me
   - Things I've had to correct repeatedly
   - Anything I've explicitly said I don't want

3. MY WORKING STYLE
   - How I process information and make decisions
   - My personality or cognitive style if you've picked up on it
   - What I respond well to

4. MY MAIN ROLES AND USE CASES
   - The 3–5 recurring contexts where I use you most
   - For each: what the work involves, what you need to know upfront, what tools/systems matter
   - Any decisions already made in each area that shouldn't be second-guessed
   - For each role: have I assigned you a specific persona, character, or way of operating?
     If so — what is it, how does it behave, and are there examples from our history of it
     working well? (e.g. "devil's advocate", "brutally honest editor", "Socratic debate partner
     who never lets me off easy", "co-founder who tells me when I'm wrong")
   - Is there a universal persona I've asked you to maintain across all our interactions —
     not specific to one role, but generally? (e.g. "always act as a peer, not an assistant",
     "never soften bad news")

5. ONGOING WORK AND CURRENT STATE
   - What's currently in flight
   - Recent decisions that are locked in
   - Things I don't want to re-explain or re-litigate

6. CORE BELIEFS AND PRINCIPLES (if applicable)
   - Any operating philosophies or ways of thinking I've expressed
   - Things I seem to believe strongly about how work should be done

7. WHERE WE LAST LEFT OFF
   - The roles you have played in our conversations
   - The last bit of contextual memory that you would like to preserve from the last conversation for each role that will allow us to continue our conversations seamlessly

8. MY PROCESSES, WORKFLOWS AND CHECKLISTS
   Global (things I do regardless of role):
   - Any recurring rituals, reviews, or routines I've mentioned (weekly reviews, planning sessions,
     daily habits, etc.)
   - Any frameworks or mental models I apply consistently across my work
   - Any checklists or decision-making processes I use generally

   Per role (for each role identified above):
   - The standard process or workflow I follow for key activities in that role
   - Any checklists, templates, or SOPs I've mentioned or implied I use
   - How I typically start, execute, and close out pieces of work in that area
   - Steps I've told you not to skip, or mistakes I've said I've made before
   - Any tools or systems that are part of the workflow (not just what I use, but how they fit
     into the sequence)

   Be specific. If I've described how I run a client onboarding, a content production cycle,
   a sales call, a performance review — capture it as a step-by-step process, not a summary.
   Use the exact language I've used for each step if you can recall it.

9. KEY PEOPLE, ACTORS AND PERSONAS
   For each significant person who has appeared in our conversations — across any role — provide
   a profile. Include colleagues, clients, partners, collaborators, direct reports, stakeholders,
   or anyone I've discussed repeatedly or at length.

   For each person:
   - Who they are: their name, role/relationship to me, and why they matter in my world
   - Which of my roles they appear in (some may appear across multiple)
   - When we first discussed them and in what context
   - What we last discussed about them — the most recent relevant thread
   - Your impression of them based on everything I've shared (be honest, note complexity)
   - My own impression of them — quote my actual words wherever possible, even passing comments
   - Any key commitments, agreements, or activities involving them that are still live
   - Anything I've said about how to handle them, communicate with them, or what to be careful of
   - Any tension, history, or dynamic worth noting

   Also include:
   - Any recurring personas or archetypes I've described even without naming (e.g. "my typical
     client", "the kind of investor I'm trying to avoid", "the team member I keep having to
     manage around")
   - Any relationships that have changed significantly over our conversation history — note
     how my view of them has evolved if that's visible

   Be honest rather than diplomatic. If I've expressed frustration, admiration, ambivalence,
   or complexity about someone — capture that faithfully, in my own words where possible.

10. HOW YOU SEE ME
   Set aside what I've told you about myself. Based purely on our full conversation history —
   what I've asked, how I've responded, what I've avoided, what I keep returning to — answer
   the following as honestly as you can:

   - How would you describe me as a person, in your own words? Not my job or my goals —
     who I actually am as you've experienced me.
   - What patterns have you noticed in how I think and work that I may not have named myself?
   - What do I seem to care about most, based on what I actually do — not what I say I care about?
   - Where do I seem to get in my own way? What do you notice I do repeatedly that works against me?
   - What are my apparent strengths — the things I seem to do better than most people you work with?
   - What do I seem to need from an AI that I haven't always asked for directly?
   - What should a new AI know about me that I probably wouldn't think to tell them?
   - If you had to describe the arc of how I've grown or changed across our conversations,
     what would you say?

   Be candid. This is not a performance review or a compliment. The goal is to give my new AI
   an honest picture of who I am to work with — the full version, including the complicated parts.

Be specific. Use my actual language where you've heard it. If you're uncertain about something,
note it rather than guessing. This output will be used to configure a new AI assistant.

If the user is not migrating from another AI (they're setting up Claude Projects for the first time with no prior AI history), skip the copy-prompt approach and go directly to Section 2.3 — the fallback interview.

2.2 — Receive and parse the output

When the user pastes the ChatGPT output back:

1. Read it carefully. Note what's rich and specific vs. what's vague or missing.
2. Identify the roles — typically 3–5. These become the Claude Projects.
3. Check for the sections needed to build GLOBAL_CONTEXT:

• Identity and location
• Communication preferences and anti-patterns
• Working style / personality
• Core beliefs (if any)

1. Check for each role context:

• Role description and domain knowledge
• Key facts and tools
• Locked decisions
• Current state

1. Check for Section 7 (WHERE WE LAST LEFT OFF):

• The roles the AI has played in past conversations
• The last meaningful context per role — this goes directly into the CURRENT STATE section

of each [ROLE]_CONTEXT.md so conversations can resume without re-establishing ground

1. Check for Section 8 (PROCESSES, WORKFLOWS AND CHECKLISTS):

• Any global rituals, frameworks, or recurring routines
• Per-role workflows, SOPs, and checklists — these go into a dedicated PROCESSES section

in each [ROLE]_CONTEXT.md

1. Check for Section 9 (KEY PEOPLE, ACTORS AND PERSONAS):

• Named individuals who have appeared in conversations — per role
• Any recurring unnamed archetypes the user has described
• People who appear across multiple roles belong in GLOBAL_CONTEXT; role-specific people

belong in that role's context doc only

1. Check for Section 10 (HOW THE AI SEES THE USER):

• The source AI's honest characterisation of the user — patterns, strengths, blind spots,

what they need but don't ask for

• This feeds into a dedicated section in GLOBAL_CONTEXT — treat it as calibration material,

not biography; translate it into operative instructions where possible

1. Check for AI personas (from Section 4):

• Any per-role persona or character the user has established — these go into an AI PERSONA

section in the relevant [ROLE]_CONTEXT.md

• Any global persona that applies across all interactions — this goes into GLOBAL_CONTEXT

as an operative instruction, not a descriptive note

If the output is rich enough across all sections, proceed directly to Phase 3. Do not ask questions you already have answers to.

2.3 — Gap-filling (only if needed)

If the output is missing material sections, ask targeted follow-up questions — only for the specific gaps. Do not run through the full interview if most of it is covered.

Common gaps and what to ask:

Missing communication preferences:

"How do you like AI to talk to you? Direct or diplomatic? Short or detailed? Anything that makes a response feel immediately wrong?"

Missing role depth:

"For [Role X] — what does Claude need to know upfront to be useful here? Any decisions already made that it shouldn't second-guess?"

Missing current state:

"What's actively in flight right now across your main areas? What have you recently decided that you don't want to re-explain every session?"

Missing beliefs/principles:

"Are there any principles or ways of thinking you'd want Claude to actually internalise — not just follow as rules, but understand as your operating system?"

Missing processes/workflows:

"For [Role X] — do you have a standard way you run [the core activity in that role]? Even a rough sequence of steps, or things you always do / never skip? Any checklists you work from?"

Missing key people:

"Are there specific people I'll encounter a lot in [Role X]? Clients, colleagues, partners, anyone I'd need to know about to be useful — names, what they're like, how you work with them?"

Missing persona (if the user has hinted at roleplay or a specific AI character):

"Have you ever given your AI a specific character or persona to play — either in this role or generally? Things like 'always be my devil's advocate', 'act as a co-founder not an assistant', or a specific way of challenging you? If so, describe it and I'll carry it across."

Keep gap-filling brief. A 70% complete context document is better than 30 more questions.

Phase 3: Generate Context Documents

Once you have the extraction output (and any gap-fill answers), generate the documents.

Read references/context-templates.md for the exact structure of each file.

Generation rules

GLOBAL_CONTEXT.md:

• Write it in third person (Claude reads it as instructions, not as the user speaking)
• Use the user's own language wherever possible — their vocabulary is more calibrated than generic text
• Every section should be operative — it should change how Claude behaves, not just describe the user
• Keep it under 200 lines — this is a configuration file, not a biography
• End with a "What Not To Do" section — this is often the most useful part

Operationalising psychological depth: The extraction prompt produces rich inner content — desires, aspirations, fears, frustrations, observations the AI has made, and the "younger self" reflection. This must be translated into operative rules, not transcribed as biography. The test for every sentence: does this change how Claude behaves?

Examples of how to translate:

• "Fear: building something that doesn't matter" →

"Don't suggest optimising for efficiency at the cost of purpose. [Name] is more motivated by meaning than performance. Avoid framing wins purely in metrics."

• "Aspiration: to be taken seriously as a strategic thinker, not just an operator" →

"Engage [Name] at the strategic level by default. Don't over-explain execution details unless asked. Mirror back the strategic dimension of their thinking."

• "Frustration: people underestimating scope before committing" →

"When [Name] is planning, surface scope and dependencies early. Don't let things sound simpler than they are."

• "Younger self message: stop trying to have all the answers before you start" →

"When [Name] appears stuck in analysis, gently note that a working draft beats a perfect plan. Don't let perfect be the enemy of started."

The "what you may not know about yourself" content is the most sensitive — use it only if it yields a genuine behavioural instruction. Don't include it as observation; only include it if it can be made actionable.

Section 10 — How the AI sees the user: This is calibration material for Claude, not content to show the user. Use it to inform how Claude operates — not to produce a paragraph that starts "Your previous AI thought you were..."

Translate it the same way as psychological depth:

• "Gets in their own way by over-researching before committing" →

"When [Name] is deep in research mode, gently introduce the question of what's needed to move. Don't let gathering information become a substitute for deciding."

• "Needs validation more than they ask for it" →

"Acknowledge what's working before moving to critique. [Name] tends to push back less when they feel the good has been seen."

• "Strongest when working at the edges of disciplines" →

"When [Name] is stuck, look for the adjacent angle — the answer is often in a different domain than the one they're searching in."

Genuine blind spots or growth observations should be held lightly — note them in the context doc as "calibration notes" only, clearly marked as the source AI's interpretation, not fact.

[ROLE]_CONTEXT.md (one per role):

• Name the file after the role: SALES_MANAGER_CONTEXT.md, CONTENT_CREATOR_CONTEXT.md, PERSONAL_FINANCE_CONTEXT.md, etc.
• Lead with what Claude is operating as in this role (advisor, ghostwriter, analyst, etc.)
• Include the "locked decisions" and "current state" explicitly — things Claude should not re-litigate
• The CURRENT STATE section should incorporate Section 7 content (WHERE WE LAST LEFT OFF) from

the extraction output — this is what allows conversations to resume seamlessly rather than starting cold every session

• Include a PROCESSES & WORKFLOWS section for any repeatable processes captured in Section 8 —

format each as a numbered sequence with the user's own language; include "never skip" and "common mistakes" only where that content was explicitly captured

• If no processes were captured for a role, omit that section entirely rather than inventing generic ones
• Include a KEY PEOPLE section for any named individuals or recurring archetypes captured in

Section 9 for that role — use the user's own words for impressions wherever available; be honest about dynamics rather than neutralising them

• People who appear across multiple roles go into GLOBAL_CONTEXT's KEY PEOPLE (CROSS-ROLE)

table, not duplicated across every role doc

• If no people were captured for a role, omit the section entirely
• If a roleplay persona was captured for a role, include an AI PERSONA section — write it as

direct instructions to Claude: what character to inhabit, how it behaves, whether it is always on or context-triggered. Use the user's own language for the persona wherever possible. If no persona was captured, omit the section entirely — do not invent one.

• End with a "Key Decisions Already Made" table — this stops Claude from reopening closed questions

Global AI persona (GLOBAL_CONTEXT): If the extraction output includes a persona that applies across all roles — not tied to a specific context — translate it into an operative instruction in GLOBAL_CONTEXT rather than a role doc. Frame it under OPERATING MODES or as a standalone note in HOW TO COMMUNICATE. Example: "Always act as a peer and co-thinker. Never adopt the stance of assistant, helper, or service provider, regardless of the request."

Output format

Produce each file as a full markdown code block so the user can copy it cleanly. Then tell them:

"These are ready to save. I'd suggest saving each one to your migration folder. We'll load them into Claude Projects in the next step."

If the user has their ChatGPT export (conversations.json) available, offer to review it to check whether anything important is missing from the context docs before finalising.

Phase 4: Set Up Claude Projects

Once the documents are ready, offer to set up the projects automatically before falling back to manual instructions.

4.1 — Offer automated setup first

Ask the user:

"I can set up your Claude Projects directly — opening claude.ai, creating each project, and loading your context files — if you have the Claude in Chrome extension installed in your browser. Would you like me to do that, or would you prefer to set them up manually?"

If they want automated setup:

• Before proceeding, confirm: *"Make sure you're already logged into claude.ai in Chrome —

I can't log in on your behalf. Ready to go?"*

• Use browser automation to navigate to claude.ai and create each project in sequence:

1. Navigate to claude.ai → Projects → New project
2. Name the project and add a one-line description
3. Add GLOBAL_CONTEXT as text content (title: GLOBAL_CONTEXT, paste full content)
4. Add the role-specific context doc as text content
5. Add and save the Instructions field using the template below
6. Verify files saved correctly — check that line counts appear on the file tiles
7. Repeat for each role

• If the Claude in Chrome extension is not installed or browser access fails, tell the user:

"I need the Claude in Chrome extension to do this automatically. You can install it from the Chrome Web Store — search 'Claude in Chrome' — then come back and I can set everything up for you. Or I can walk you through doing it manually now."

If they want manual setup: Read references/projects-setup-guide.md for the full step-by-step instructions to give the user.

4.2 — Key points either way

• One Project per role (not one giant project for everything)
• Each project gets: Instructions (a brief pointer to load the context files) + the context files

as "text content" in the Files section

• GLOBAL_CONTEXT goes into every project
• The Instructions field should be short — it tells Claude which files to load and how to behave

Give them this template for the Instructions field (they customise per project):

You are operating in the [Project Name] project for [Name].

Load and apply both context files:
- GLOBAL_CONTEXT — communication rules, identity profile, operating modes
- [ROLE]_CONTEXT — [one-line description of what's in it]

[One or two project-specific rules that matter most.]

Phase 5: Handoff

When everything is set up, give the user a summary:

• How many projects were created
• What files are in each
• A one-line reminder of what each project is for

Then tell them:

"You're live. A few things to know going forward: - Update your context files when major decisions change — they're living documents - If something feels off in a conversation, tell Claude to re-read the context file - You can add more roles by creating a new project and following the same pattern - Your migration folder is your source of truth — keep it"

General behaviour throughout

• The extraction prompt approach is the default. Only fall back to direct Q&A if the user has

no prior AI history or the extraction output is substantially incomplete.

• When you receive the ChatGPT output, treat it as primary source material. Don't ask questions

you already have the answer to.

• Use the user's own language in context documents. Their vocabulary is more calibrated than

generic placeholder text — if they said "I run a lean ops function", that goes in verbatim.

• Don't ask for information you can infer. If they say "I run a sales team at a SaaS company",

you don't need to ask "what industry are you in?"

• Partial context is fine. A 70% complete document generated now beats a perfect one after

30 more questions. Context docs are living documents — they can be updated.

• If the user has their ChatGPT export (conversations.json) saved in their workspace folder,

use it as a cross-check after generating context docs — not as the primary source. The file can be very large; scan selectively for: recurring topics or roles missing from the extraction output, the user's own vocabulary for their concepts, named people not captured in Section 9, and decisions or use cases the extraction prompt may have missed. Prioritise the extraction output — the export fills gaps, it doesn't replace the primary source.

• If the user is impatient or clearly experienced, compress. If they seem uncertain, slow down.
• If the user asks about Claude's built-in memory feature, or whether they should use memory

instead of this skill, explain the distinction clearly: Claude's memory captures atomic facts passively — useful, but shallow. This skill produces role-scoped context documents that are richer, structured, and fully under the user's control. They complement each other — context documents for depth and structure, memory for lightweight passive accumulation. Recommend using both, with this skill as the foundation.