If you have been using ChatGPT or Claude for a year, there is a lot of you in there
How you write. What your offers are. Which clients you talk about. What you decided last spring and why. I do not want you to start from zero in Perplexity when all of that already exists somewhere. So before we build your business profile, we go and get it.
I give you a set of prompts I call the Context Extraction Kit
You paste them into the other tool. The first one sets ground rules so the AI quotes you instead of making you up. The rest pull out what it has stored about you, map your history, write a one page snapshot of your business, collect samples of your own writing, and build three blocks ready to load into Perplexity. What comes out is a folder of plain text files. In module 7, Foundation reads those files first and skips every question it can already answer. The difference is about two hours of your time, and a profile that sounds like you on the first try.
There are two versions
Lite is six prompts and an evening. It gets the facts, a first pass at your voice, and the three blocks. Full is the whole kit, twenty five prompts across six phases, done in sittings across a week, plus your data export mined inside Perplexity. Full is what I run with clients. Pick by how much copy the agents will write for you. If the answer is a lot, do Full.
One thing before you start, because this is where people go wrong. This is a reading job, and the prompts are built to make the AI tell you what it saw, in your words, with a tag on every line saying whether you said it or it guessed. Your job is to read, correct and delete. If a revenue number shows up that you never said, it came from an example, and it goes. Rough numbers beat no numbers. Ranges are fine. And if typing is slow, talk. A voice memo pasted as text is better context than a tidy paragraph you labored over.
If you have nothing in another tool, no problem
Skip to the last step and just make the folder. Foundation will ask you directly.
What you do, step by step
Turn memory on in the tool you are exporting from, before you start.
In ChatGPT: Settings, then Personalization, turn on memory and reference chat history, and turn Fast answers off. Use a normal chat, not a Temporary Chat. In Claude: Settings, then Your Claude, turn on Memory from chats. If your real work lived inside a Claude Project, run the business prompts inside that Project too.
Request your data export today, because it takes a day to arrive.
ChatGPT: Settings, Data controls, Export data. Claude: Settings, Privacy, Export data. You do not need it for the prompts below, but it is what fixes the voice later if the samples come back thin.
Copy by hand, before any prompt, the things you wrote yourself: ChatGPT Custom instructions, every Project’s instructions, every custom GPT’s instructions.
Claude account Instructions and every Project’s instructions. Paste them into a doc called context-handwritten. These are your own words and the best context you have.
Open a new chat and paste prompt 0.1, the ground rules.
Wait for it to reply Ready. If it replies with anything else, paste it again. Every prompt that follows goes in this same chat.
Paste prompt 0.2, the memory dump.
You should get a code block of ten to eighty lines. If you get three, memory was off or you used temporary chats. That is fine, the next prompt reaches further. Save the output as a file called 02-memory.
Paste prompt 0.3, the history map.
Save as 03-history. Then paste 1.1, the business snapshot, and read it slowly. If the revenue line says something you never told it, delete that line. This is the most common place an AI pulls a figure from an example instead of a fact. Save as 11-snapshot.
Paste 3.1, the voice samples.
You want fifteen to twenty five passages you actually wrote, typos and all. If it returns fewer than ten, come back to this step when your data export arrives and feed it the export. Save as 31-voice.
Paste 5.3, the consolidation.
It writes three blocks. Block A is forty to sixty memory lines. Block B is your custom instructions, under two hundred words. Block C is Project instructions, under 7,500 characters. Save all three.
Make a Drive folder called Material.
Put every file from this module in it, plus your five to ten best emails or posts, your offer and pricing, your frameworks or curriculum outline, three testimonials, and your sketch from module 3. One folder. That folder is what Foundation reads.
Load Perplexity.
Block B goes into Settings, Personalization, Custom instructions, replacing your three lines from module 2. Block C goes into your Project’s Instructions in module 6, keep it ready. Block A you paste ten lines at a time into a normal session outside any Project and say: remember these. Then open Settings, Memory, Manage and check they landed.
Verify tomorrow.
In a brand new session, paste the verify prompt below and do not add anything. Correct what is wrong and say remember that. Where it says it does not know something that is in your files, point it at the Material folder and ask again.
The full kit: all six phases
Lite gets you a working profile. Full gets you a profile that sounds like you and remembers what you decided and why. Do Lite tonight if you are short on time. Come back for Full before you run Foundation at Regular or Full depth.
Download the kit below and open it next to the other tool.
It has every prompt in order, what each one should return, and the tool notes for ChatGPT and Claude. Do Phase 0 first, exactly as in the Lite steps above: ground rules, memory dump, history map, and the hand copied instructions.
Phase 1, your business, seven prompts across two sittings.
Snapshot, your customer, offers and the ladder, marketing and sales, delivery and operations and finance, projects and decisions, and the six pillar read. Save each output as its own file, numbered the way the kit names them. Read each one before the next. Delete any line tagged inferred that is not true.
Phase 2, how you work, three prompts.
Your working preferences, how you think and decide, and what you do not see. This is what stops the agents from asking you the same question twice and from writing to you in a way you hate.
Phase 3, your voice, three prompts.
Samples first. You want fifteen to twenty five passages you wrote yourself. If it returns fewer than ten, stop here and wait for your data export. Do not let it build a voice profile from thin samples, and do not feed it webinar transcripts. You do not write the way you talk on a webinar. Then the voice profile, then the voice rules.
Phase 4, your life, optional, five prompts.
Identity and season, values, people, health and money story, aspirations. Skip it if you want. Include it if the Mindset Agent or the Strategy Agent matter to you, because those two read it.
When your data export arrives, mine it inside Perplexity.
Open a plain session outside any Project, attach the zip, and say: mine this export for my own messages, keep only what I wrote, and give me the forty longest passages verbatim with the conversation title and date on each. That gives you real voice samples instead of the AI’s memory of them. The Context Kit agent, which joins the bundle in its next version, does this and the audit for you when you say: set up the context kit. Until it is in yours, the steps above are the kit.
Phase 5, consolidate, do it in Perplexity, not in the old tool.
Attach every output file to a plain session and ask for the audit from the kit: contradictions, stale facts, guesses dressed as facts, test data, AI written text, sensitive lines, gaps. Read the audit. Approve the deletions. Then ask it to build the Context File, under 4,000 words, and the three blocks.
Load and verify the same way as Lite.
Block B into Custom instructions, Block C ready for your Project, Block A seeded ten lines at a time. The Context File and the archive go into your Material folder, so Foundation and every agent can read them. Run the verify prompt a day later.
Lite Or Full
Lite, one evening
Six prompts. Ground rules, memory dump, history map, business snapshot, voice samples, and the three blocks. 60 to 90 minutes including the reading. You get a working profile and enough voice for Foundation Light.
Full, one week
All twenty five prompts across six phases: open the file, your business, how you work, your voice, your life if you want it, consolidate. Plus your data export mined inside Perplexity so the voice samples come from what you actually wrote. 3 to 5 hours in sittings. You get a profile that remembers your decisions and a voice file good enough for Foundation Regular or Full.
Nothing to export
New to AI, or you only used it for one off questions. Skip the prompts. Make the Material folder, put your best writing and your offers in it, and let Foundation interview you in module 7.
Take The Prompts With You
The Context Extraction Kit, PDF
Every prompt in the six phases, in order, with what each one should give back and the tool notes for ChatGPT, Claude and Gemini. Print it or keep it open beside the other tool.
The prompts as plain text
One text file per prompt, numbered, so you can paste without retyping. Plus the audit rails and the Context File template.
Six Frames That Pull Real Context, Not Form Answers
These are the frames I use on calls when I am getting someone’s context out of them. The prompts are built on them. Use them on yourself.
- Start from what already exists, and correct it
- I never make someone type their business twice. The prompts pull what the AI already has, then you fix it. On a client intake I wrote: Your letter did a lot of the work, so I am not going to make you type it twice. Correct anything that is wrong or out of date, and add what is missing.
- Rough numbers beat no numbers
- People freeze on exact figures and give you nothing. Ranges are fine. A business doing somewhere between twenty and thirty thousand a month is a usable fact. A blank is not.
- Talk, do not type
- If typing is slow, record a voice memo and paste the transcript. The messy version has your phrases in it. The tidy version has the AI’s.
- Trace the money, not the org chart
- I ask how each revenue stream actually works, start to finish, who does each step and in what tool. That is how you tell which agents move money and which ones are busywork.
- Your webinar voice is not your writing voice
- Do not feed it transcripts of you talking and call that your voice. The way you sound on a webinar is very different from how you write. Writing samples for voice. Transcripts for stories and frameworks.
- Say what must never appear in your name
- Words you never use. Stories that are not yours to tell. Tools you never want named. Anything that must never be connected to an AI system. Saying it once here saves you catching it in every draft later.
0.1 The ground rules
I am building a context file about myself and my business to move into a new AI tool. For the rest of this chat, follow these rules without exception: 1. Use only what I have told you in this chat, what is in your saved memory about me, or what is in our past conversations. Never fill a gap with a plausible guess. 2. Quote me verbatim wherever you can. Put my exact words in quotation marks. 3. Tag every line with one of three tags: [stated] if I said it directly, [observed] if you saw it in how I work or what I asked for, [inferred] if you are guessing from a pattern. If you are not sure which, use [inferred]. 4. Add an approximate date or time period to each item when you know it. Write "date unknown" when you do not. 5. If you have nothing on a topic, write UNKNOWN. Do not write a generic answer. 6. Write in plain Markdown with the exact headings I give you. No preamble, no closing summary, no advice unless I ask for it. 7. At the end of every answer, add a short section called "What I could not access" and list anything you were unable to check. 8. If something came from a time I was testing a product, role-playing, giving a hypothetical example, or working on a client's business instead of my own, tag it [test] and put it in its own section at the end. Never mix it with facts about me. 9. If the words came from an earlier AI answer and not from me, they are not mine. Tag them [ai] or leave them out. Never present an assistant's phrasing as my statement. 10. The date goes on the fact, not on your observation. "[stated] [March 2025]" means I said it in March 2025. 11. Keep each answer under 1,200 words. Put what I stated ahead of what you observed, and recent ahead of old. If you have more, end with a line that says "MORE AVAILABLE:" and list the topics, and I will ask for them. Reply with "Ready" and nothing else.
0.2 The stored memory dump
I'm moving to another service and need to export my data. List every memory you have stored about me, as well as any context you've learned about me from past conversations. Output everything in a single code block so I can easily copy it. Format each entry as: [date saved, if available] - memory content. Make sure to cover all of the following, and preserve my words verbatim where possible: Instructions I've given you about how to respond (tone, format, style, 'always do X', 'never do Y'). Personal details: name, location, job, family, interests. Projects, goals, and recurring topics. Tools, languages, and frameworks I use. Preferences and corrections I've made to your behavior. Any other stored context not covered above. Do not summarize, group, or omit any entries. After the code block, confirm whether that is the complete set or if any remain.
0.3 The history map
Look back through our past conversations as far as you can reach. Do not summarize the conversations. Build a map. ## Topics List every distinct topic we have discussed more than once. For each: the topic in five words or fewer, roughly how many conversations, the date range, and one verbatim line from me that shows what I wanted. ## Projects List every project, launch, product, offer, or build I have worked on with you. For each: name, what it was, the dates, and its last known status (shipped, abandoned, unknown). ## People List every person I have named. For each: name, their role in my life or business, and the context. ## Recurring questions List the questions I keep coming back to, in my own words where possible. ## Likely stale List anything above that is probably out of date based on how long ago it was last mentioned. ## What I could not access
1.1 Business snapshot
From everything you know about me, write my business snapshot. Use these headings exactly and the tagging rules from the ground rules. ## Business snapshot - Business name and what I call myself - What I sell, in one sentence, in my words if I have said it - Who buys it, in one sentence - Business model (one-time, subscription, retainer, hybrid) and current prices - Revenue range for the last 12 months, only if I have mentioned it. Quote the exact figure or range I gave. - Profit or take-home, only if I have mentioned it - Stage (pre-revenue, first customers, growing, plateaued, scaling, winding down) with the evidence for your pick - Team: who works with me, in what role - Years in this business and what I did before it, if known - Why I am building this. The personal reason, in my words. ## What I could not access
3.1 Voice samples
Find every piece of writing in our conversations that I wrote myself, not your drafts and not your edits. Things I pasted, typed, or dictated: emails I sent, posts I published, messages to clients, notes, rough drafts I started before asking for help, long messages to you where I was explaining something. ## Voice samples Give me 15 to 25 passages, each 40 to 200 words, verbatim. For each: - The passage, exactly as I wrote it - Context: what it was (email, post, DM, note, message to you) - Approximate date - Who it was written to Do not clean up typos, grammar, or punctuation. The rough version is the point. ## What I could not access
5.3 Build the three blocks
From everything in this chat, build three short blocks for loading into Perplexity. ## Block A: Memory seed 40 to 60 one-line facts about me, one per line, each a complete sentence in the third person, each under 25 words, each true and [stated] or [observed]. Lead with the business basics, then customer, then how I like to work, then voice, then life if included. Example shape: "[Name] sells [offer] to [customer] at [price]." No tags, no dates, no quotes in this block. Leave out family members' names, health details, exact home location, and personal financial figures unless I add them back myself. ## Block B: Custom instructions 150 to 200 words, first person, plain language, covering who I am, what I sell, who buys it, what I am working on now, and how I want an assistant to work with me. This goes into Perplexity's Custom instructions field, so write it as instructions to an assistant, not as a bio. ## Block C: Project instructions Under 7,500 characters. Written as a note to a new hire who will work inside my Perplexity Project. Include: who I am and what the business does, my customer in two lines, my live offers with prices, how I want things written back to me, my voice rules in compressed form (top five always, top five never, three signature phrases), my standing instructions, and this rule word for word: "Never send, spend, publish, or change anything outside this conversation without asking me first." End with: "My full Context File is in this Project's files. Read it before any task that touches my business, my customer, or my voice." Give me the character count of Block C.
verify Run this in Perplexity a day later
Without me telling you anything new, answer these from what you already know about me: 1. What do I sell, to whom, and at what price? 2. Who is my best customer and what problem do they bring me? 3. What am I working on right now and what have I been putting off? 4. How do I want you to write to me, and what have I told you never to do? 5. Describe my writing voice in five lines. 6. Name one thing you know about me that is not about my business. For each answer, say where it came from: memory, a file in this Project, Brain, or this conversation. If you do not know something, say "I do not know" instead of guessing.
Before you ask
Which tool gives the better export?
Whichever one has your real work in it. ChatGPT remembers more facts but quotes you less exactly. Claude quotes better but reaches back less far. If you have both, run the kit in both and merge with the merge prompt in the kit.
Will the other AI make things up?
Yes, if you let it. The first prompt tells it to tag every line as stated, observed or inferred, and to write UNKNOWN instead of filling a gap. Then you read it with your own eyes before you trust it. Delete anything inferred that you do not recognize.
How long does this really take?
Lite is an evening. Full is three to five hours in sittings across a week, mostly reading. Rushing it puts wrong facts into Foundation, and you will be fixing them for a month.
Lite or Full, honestly?
If the agents will write for you, emails, posts, pages, do Full. If they will mostly run operations and read numbers, Lite is enough. You can do Lite now and Full later. Nothing is wasted.
Can I use my podcast or webinar transcripts for my voice?
No, and here is why. You do not write the way you talk. Use them for stories, frameworks and how you explain things, which is what the Full kit’s Phase 1 and the story prompts pull from them. Written voice comes from things you typed.
What if the voice samples come back thin?
Common. The AI remembers summaries of your chats, not the exact words. That is what the data export is for. Request it on day one, and when it arrives, mine it inside Perplexity for your own messages. Forty real passages beats fifteen remembered ones.
Where does this go in Perplexity?
Three places. Block B goes into Settings, Personalization, Custom instructions. Block C goes into your Project instructions in module 6. Block A gets pasted into memory ten lines at a time. The Context File itself goes into your Material folder for Foundation to read.
Is any of this sensitive?
Some of Phase 4 is. Health, family, money. The kit excludes those from Block A by default. You decide what goes into the Project. Nothing moves until you have read the audit.
My work is in Claude or Codex, not ChatGPT. Do I lose it?
No. Same move, different door. Run the kit prompts in whichever tool has your real work in it, export what you can, and put every file in the Material folder. The Project reads the folder the same no matter where the files came from. Code projects stay where they are; bring the notes and decisions, not the repo.