The interview
The hour that turns a blank machine into an assistant that knows your world, and what it asks about.
You will be able to give your assistant a brain of its own by answering its questions, and pick up where you stopped.
This is the part that makes the difference between an assistant and a chatbot with your logo on it, and it is the part people put off.
A new mineral knows nothing. The interview is how it comes to know your world: what you are trying to do, who is in it, how you work, what to never touch. It takes about an hour, it does not have to happen in one sitting, and every other page in these docs assumes it has happened.
Starting it#
Open your mineral in Claude Code (the app's Help page shows the connection; for a folder on this computer, Open folder and pick it), type this, and press Enter:
/onboardThat is the whole ceremony. If you are partway through from an earlier sitting, the same command resumes it; you never have to remember where you were.
The Terminal tab in the app also accepts it, and it is fine for a quick answer, but the interview is an hour of conversation and Claude Code is the better room for it: a bigger window, your whole brain in context, and it is where you will work day to day.
The first question is what to call it#
Before anything else, it asks what you want to call it, and writes that down immediately rather than at the end.
That is deliberate. Naming something you are about to talk to for an hour is a natural thing to be asked. Naming it in a setup form is naming a server, which is a colder and different act. And because it is written on the first turn, someone who walks away mid-interview comes back to something with a name rather than to "your assistant".
One name, everywhere. The mineral and the assistant share it: the sidebar, your pebble page and the greeting all say the same word, and the Rename button changes that one thing. Renaming never touches the address, the connection or the keys.
If your mineral already has a name, it will not ask. If you have no preference, it keeps the one it has. It will not invent one for you, and it will not offer you a list to pick from: a name you did not choose and do not recognise is exactly the confusion this question exists to prevent.
The eight layers#
The interview walks eight layers, in order, and does not move on until the current one is genuinely covered rather than merely touched.
| Layer | What it is really asking | |
|---|---|---|
| 1 | North star | the long-term why, and what winning looks like in five to ten years |
| 2 | Philosophy | your values, what you optimise for, what you will not compromise, how you decide, and your anti-values: what you will not do |
| 3 | Self | your story, what energises and drains you, how you handle pressure, the health and family context that affects work, and your actual work: what you offer, to whom, at what price, the key numbers, your seasons, your bottlenecks |
| 4 | Network | the people. Customers, suppliers, team, mentors, family. Who they are, what they need, who owes whom, what is going cold |
| 5 | Past | what you have built, the war stories, the lessons |
| 6 | Goals | this quarter and this year, with success criteria, and your anti-goals |
| 7 | Tasks | what you are actually working on right now |
| 8 | Workflow | your rhythm, how you like to be nudged, what to handle and what to never touch, your tools, what you wish was automated, your biggest time-sinks, how you write and speak, and your hard limits |
None of this stays abstract. Every layer becomes a literal page you can open: one page per layer, under _layers in your brain, plus a page per person you mentioned. Each person's page carries a tier, a note of how load-bearing that relationship is, and the tier is not ceremony: the follow-up sweep later scans those same pages for people left in limbo and weights the load-bearing relationships first. All of it is visible on the Brain page afterwards.
Layer 8 also captures real samples of your writing, which is what lets your assistant draft as you rather than as a language model. How that voice keeps improving after the interview is teach it your voice.
How it actually feels#
One question per turn. Not a form, not a wall of questions. It asks, you answer, it follows the thread you opened rather than reading down a checklist. This is the interview step of the CRIT method, run at the scale of a whole brain; you can ask for the same thing on any task.
It probes the thin bits. If an answer skims a layer, the next question goes at the thinnest point rather than moving on politely.
It stops when a layer is covered, not when the questions run out. And covered is assessed, not assumed: each layer has a checklist behind it, and every turn your answers so far are checked against it. A point counts when it has been meaningfully addressed, not merely touched.
You can stop whenever you like#
There is no pause button because none is needed: just stop answering, close the window if you like. Progress is written to your mineral after every single turn, so nothing you have said is ever lost, and typing /onboard next time resumes exactly where it stopped, days later if need be. A resumed interview greets you by the name you chose, and having a name it never asks again.
The mechanics are plain enough to state. The interview moves through four phases (interview, synthesis, review, done), and each layer carries its own status. That state is read at the start and written at the end of every turn, which is exactly why closing the window loses nothing: there is never a turn's worth of progress held only in the conversation. "Done" is not just a feeling, either; it is the signal other parts of the product read to know this brain is seeded.
That is the honest way to do it: an hour is a lot to find in one piece, and three twenty-minute sittings produce a better brain than one exhausted hour. The Onboarding card on your Overview counts the layers as they land, so you can always see how far in you are.
What happens at the end#
Two passes, not one.
Synthesis. Your answers become proper pages: written up, cross-linked, with a page per person you mentioned. Not a transcript. Your verbatim answers are kept separately during capture, and the writing-up happens afterwards so nothing is pre-polished into blandness while you are still talking.
Review. It walks you back through each layer in plain language: here is what I captured about your north star, is that right? You correct it, and the corrections stick.
Only then is the brain seeded. You can see the result on the Brain page, and the count on your Overview reads eight of eight.
What it will not do#
It will not invent things about you. If something is undecided, it records that it is undecided rather than filling the gap with something plausible.
It will not surface the private things casually. Layers 3 and 8 collect real context: health, pressure, hard boundaries. That is stored and flagged, not sprinkled into a morning brief.
It will not finish without you. The review pass needs your corrections. A brain nobody checked is a brain nobody should trust.
The wider list, for the assistant as a whole rather than the interview, is what it will not do.