How to Set Up a New AI Assistant in 10 Minutes (2026)

Open a new AI tool and it hands you the same thing every time. A blinking box, no idea who you are, and a quiet invitation to explain yourself from the beginning.

Most people accept the invitation. They start typing background into the first chat, then more of it into the second, and a week later the new tool knows roughly what the old one knew, spread across conversations nobody can find again. That's not setup. That's re-teaching, and it's the reason switching tools feels expensive when it shouldn't.

How do you set up a new AI assistant in 10 minutes?

Paste the context file you already own into wherever the tool keeps standing instructions, add one line telling it how to treat you, then run a real task instead of a test question. Correct the first thing it gets wrong, save that correction back into your file, and stop. Ten minutes, and the tool is doing most of them.

The whole method assumes one thing: that your context lives in a document you control rather than inside the last app you used. If it does, everything below is mechanical. If it doesn't, the last section covers where to start.

What travels with you, and what you leave behind

Not everything you built in the old tool is worth moving, and some of it can't move at all.

What you're tempted to bring	Bring it?	Why
A markdown file with your role, situation, and preferences	Yes	Reads identically in every tool, including ones not released yet
Your chat history export	No	It's a transcript, and nothing ingests a transcript as context
Saved memories from the old tool	Can't	Proprietary storage with no export another product can read
Old custom instructions	Copy once	Useful raw material, but maintaining three drifting copies is the trap
Your last twenty prompts	No	Prompts are technique. Context is the thing technique operates on

The middle rows are where the ten minutes usually goes wrong. Chat history feels like the valuable part because it's the part you spent time on, but a context anchor does in one page what four hundred exported messages can't do at all.

Minutes one and two: paste the file

Every major tool has a place for text that persists beyond a single conversation. The names differ, the mechanics barely do.

Where it goes	Good for	Watch for
Custom instructions	A baseline that follows you into every new chat in that tool	Character limits, so this holds the compressed version
A project or persistent workspace	Long-running work where the full files should stay in scope	Scoped to one tool, so the setup itself doesn't travel
The top of the conversation	One-off sessions, or any tool without the other two	Manual every time, and easy to forget under deadline

Pick by how long you expect to use the thing. Trying a tool for an afternoon? Paste at the top and move on. Planning to live in it? Put the full files in a workspace. The tool-by-tool mechanics are slightly different, and none of the differences are worth more than a minute of your attention.

One rule holds across all three: paste, don't retype. Every rewrite creates a version, and versions are how you end up with three documents that quietly disagree about what you do for a living.

Minute three: one line of calibration

Add a sentence about how you want to be treated. Not what you do, how you want to be answered.

Argue with me before you agree with me. Short answers unless I ask for the long one. If I'm asking how to do something I shouldn't be doing at all, say that first, then answer anyway.

It's one sentence, and it's the one almost everyone skips. Context tells a model what's true about you. Calibration tells it what kind of help you actually want, which is the difference between a competent answer and a useful one.

Minutes four to eight: give it real work

The temptation here is to test it. Ask something you already know the answer to, see whether it sounds smart, grade the personality.

Don't. Test questions measure charm. Pull something from your actual queue instead — the email you've been avoiding, the plan that needs a second opinion, the decision with two defensible options. Real work is the only thing that reveals whether the context is loaded or just present.

Watch for one specific signal: does the answer reflect your situation, or the average situation? A generic response to a specific question means the file is sitting in the session without being used, and the fix is usually a pointer rather than more text. One line — read the attached context before answering anything involving a trade-off — gets a model to actually consult what you gave it.

Minute nine: correct exactly one thing

Find the first place the output missed and say what you wanted, plainly. One correction. Not a list.

The instinct is to fix everything at once, which turns into rewriting the output yourself and teaches the tool nothing. One clear correction shows you where the briefing was thin, and that's the information you came for.

Minute ten: write the correction back

Take the sentence you just used to correct it and add it to your file.

This is the step that makes the previous nine minutes permanent. Fix it in the chat and you've fixed one conversation. Fix it in the file and every tool you open from here starts correct, including the one you'll be evaluating a year from now. It's the same discipline that keeps context from going stale, applied at the moment the gap is most obvious.

Then stop. Ten minutes is the measure of whether your context was portable in the first place. A setup that runs an hour usually means it wasn't, and you're rebuilding it live in a chat window.

What if you don't have a file yet?

Then the ten minutes buys you less, and the honest answer is that your first setup will take longer than your second.

Start with the smallest thing that works. Three paragraphs: what you do, what you're working toward this quarter, and how you want to be answered. That's enough to make the first real task go better than it would have, and it's enough to prove the mechanic before you invest in anything larger. The six anchors are where it goes from there, one durable question at a time.

Why the ten minutes is the whole argument

Portable context sounds like insurance against something that will probably never happen. A tool shuts down, a model gets worse, a price changes.

But the everyday payoff is smaller and far more common than that. A colleague recommends something over coffee. A new model launches on a Tuesday and you want to know whether it's better at the one thing you actually use AI for. Either of those is a ten-minute door if your context is a file, and a week-long project if it lives inside a product you don't control.

So open the tool you've been meaning to try. Paste the file, add the line, hand it something real, and see what you get before lunch.

Start with the free Personal Constitution anchor if you don't have a file to paste yet — it asks the questions so you don't have to face the blank page.