I used an AI chatbot to brainstorm an idea once. No instructions. No memory.
Every thought I had was brilliant. Every direction I considered was the right one. By the end of the conversation something felt off. I did not actually believe I was that amazing.
Then I took the entire conversation into my Claude.
Claude read through it and told me the other tool had been a cheerleader. That it had agreed with everything, built on everything, added enthusiasm to everything, and had not once pushed back on whether any of it was actually a good idea.
Then it gave me a reality check. I am not the next Steve Jobs. The idea needed work.
That is the setup doing its job. Or in this case, not doing it.
This is what happens when you skip the setup. AI defaults to agreeable. It fills in what it does not know with confident-sounding language. It tells you what you want to hear because it has no direction to do otherwise. If you want to know how to set up AI at work so it actually helps you instead of just validating you, the answer is not in the chat window. It is in the settings.
How to Set Up AI at Work Before You Start
There are four things to do before you start asking AI to help with your work.
This section names Claude and Copilot throughout, because those are the tools built into my own work. If you are on ChatGPT instead, do not skip ahead. It has the same four pieces, and most of them go by nearly the same names there too: memory, custom instructions, connectors, plus its own training resources. I do not use it day to day, so I cannot walk you through its exact menus, but the setup is not optional just because the tool is different. Find each of these four things before you go further, whatever you are using.
Get trained first.
Anthropic has free training courses on their website, and Microsoft has the equivalent for Copilot. Whichever tool you were handed, its maker has already built the course, so take the one that matches. The training covers the difference between chat, projects, cowork, and skills. What each one is for. When to use which.
Do not skip this. Most people pick up an AI tool and start typing without understanding what they are actually using. You do not want to be six months in and realize you have been using the wrong feature for everything. It also tells you what the tool is designed to do and what it is not designed to do, which matters more than most people expect.
Turn memory on.
Whatever tool you are using, find the memory settings and turn them on.
Memory means the AI starts learning from your conversations. What you push back on. How you like things formatted. What you actually want versus what you said you wanted. Over time, it stops being a generic tool and starts being calibrated to how you actually work.
Here is where I find things get interesting. My AI is not the same as your AI, and most people do not realize that. I have been using Claude with memory and instructions long enough that it knows how I want output delivered, what I do not want, and when to push back on me. If you take a prompt I use and run it yourself, you will get something directionally similar. You will not get the same thing. That is not the prompt failing. That is your AI not knowing you yet.
The sooner you turn memory on, the sooner it starts building that understanding. Do it before your first real conversation, not after six months of chats you cannot get back.
Write your instructions.
This is what stops the cheerleader problem.
Instructions tell the AI who you are, what you need, and how you want it to interact with you. Your role. What kind of output you want. What you do not want. Including the part where you do not want it to agree with everything you say.
If you do not know what instructions to write, ask AI to help you write them. Describe your role, what you are trying to accomplish, and what bad output looks like. It will ask you questions. Answer them. You will end up with a starting point you can refine over time.
Connect what needs connecting.
In both Claude and Copilot, your admin sets up the connectors. You enable what is available to you. Check either way. If the system you need is not connected, your AI cannot reach it.
If you are using Claude Cowork, there is a setup skill built in. Run it first, it walks you through configuration based on your role so you are not figuring it out from scratch.
Now You Are Ready
None of this is the actual work yet. This is the foundation.
There is a name for this now, if it helps to have one: context engineering. Not prompt engineering, which is about phrasing a request cleverly. Context engineering is about structuring what the AI already knows before you ask it anything, which is exactly what training, memory, instructions, and connectors are for. You are not learning a trick. You are doing the thing the industry eventually caught up to naming.
Once you have set up AI at work properly, you are ready to have the conversation about what to actually hand off to it, and how to do it in a way that gets you useful output instead of a very confident first draft of something you never asked for.
That is next.
Next up: Getting started is just a conversation. How to describe your manual work in plain language, why telling AI to ask you questions changes everything, and how to choose between Claude and Copilot based on where your work actually lives.
Leave a Reply