You build an AI assistant by giving an existing AI three things: a fixed role, clear rules, and your own knowledge attached to it. You can do this without any coding, most commonly in a Claude Project, a ChatGPT Custom GPT, or an agent inside Microsoft Copilot. In all three cases, the assistant stays with text: it doesn't pick its own tools, and it doesn't make any decision beyond a draft.
I run my own business with a workforce of eight AI employees built on Claude, and one of them, Sonja, is an assistant who drafts my emails in my own voice. What an AI employee is at its core, and how it differs from a plain assistant, is collected at AI employees.
The broader concept behind it, giving an existing AI your own knowledge instead of training a model, is covered in more depth in Build Your Own AI Without Coding. What follows is the build itself, on the three platforms that come up most often.
Assistant, chatbot, or agent: where an assistant sits
An assistant is neither a chatbot without memory nor an agent that operates its own tools. It sits right in between: a fixed role, attached knowledge, but the result is always a text draft you read before anything happens. The full breakdown between chatbot, assistant, and agent, with a comparison table, is in AI Agent, Chatbot, or Assistant?. For this post, the short version is enough: a personal AI assistant writes and organizes for you, it doesn't act on its own.
Three paths to building an AI assistant
Claude Project. You set up a project, give it a fixed role in a system prompt, and attach your own documents as knowledge. According to Anthropic's help center, Claude can then work within its own chat history and your context in that project (support.claude.com, "What are projects?"). The assistant then answers every new request within that role and knowledge base, without you re-explaining the context each time.
Custom GPT in ChatGPT. You describe the role in your own words, set fixed instructions, and upload knowledge files the GPT can cite from. Building your own Custom GPT in ChatGPT usually requires a paid account, a free account is typically enough only to use GPTs other people have published.
Agent in Microsoft Copilot. If you already work inside Microsoft 365, you can set up your own Copilot agent there, with a role, instructions, and knowledge sources drawn from your files or SharePoint. The advantage is direct access to documents you already have, the downside is a tighter bond to that ecosystem.
| Path |
Attaching knowledge |
Strength |
Limit |
| Claude Project |
upload your own documents |
large context window, pricing at Claude Pricing and Plans Explained |
stays with text, no independent tool access |
| ChatGPT Custom GPT |
upload knowledge files |
quick to share and reuse |
building your own GPT usually needs a paid account |
| Copilot agent |
files and SharePoint as source |
direct access to existing company documents |
tightly bound to the Microsoft ecosystem |
Which path fits depends less on the technology than on your day to day: where your documents already live, which account you already have, and which surface you write the most text in.
What every assistant needs, no matter the platform
Three building blocks turn an empty window into an assistant, whichever platform you pick. The role tells it who it is for you and what it's responsible for, stated concretely instead of generally: not "help me with emails", but which emails, in which tone, toward which goal. The rules define what it may decide on its own and what it can't, for example that it never makes commitments or quotes prices without your review. The knowledge, finally, is the set of documents it should actually cite from: past emails as templates, your price list, your service catalog, but never sensitive data that has no business being there.
Getting the order right matters most. An assistant with good knowledge but an unclear role sounds informed but has no recognizable voice of its own. An assistant with a clear role but no knowledge sounds like you but often gets the substance wrong.
Sonja: an assistant with a voice, not a chatbot
Here's what that looks like in my own business. Sonja is my assistant for email, not a chatbot I open when I need it, but a fixed role with her own personnel file. Her input is a new message in my inbox. She knows my tone from past emails, my typical phrasing, and the rules for what she never decides on her own, like commitments or appointments without my review.
From an incoming message, she produces a finished reply draft in my own voice, not in a generic AI tone. The result of her work is always a draft waiting on me, never a sent email. I read it, check whether the tone and the content are right, adjust a phrase if needed, and decide whether it goes out as is. That last step, the approval, is the real difference between an assistant with a voice and a plain chatbot: a chatbot gives you an answer to one question, Sonja gives you a draft that sits inside an ongoing thread built from many earlier emails.
Test it and correct it before you trust it
The first draft from a new assistant is rarely perfect, on any of the three platforms. That's normal, and it's not a reason to rewrite the role from scratch. Instead, give concrete feedback on individual answers: "too formal", "this detail doesn't belong in a first reply", "we don't address this case that way". Add every correction to the role or the rules in writing, in the text the assistant sees on every request, not just as a spoken note.
After a few test rounds with real but low-stakes cases, you'll see whether the assistant holds up. Only then is it worth using it for cases where a wrong tone or a wrong fact actually costs you something.
From assistant to AI employee
An assistant stays with text and waits for you. Once you want the same role to research independently, use several tools, or do more than answer in text, it becomes an agent with its own personnel file, in other words an AI employee. The full path there, from picking a task to permanent training, is in Creating an AI Employee: How to Start. You don't have to take that step right away: many tasks, like Sonja's email drafts, stay with an assistant for good, because a text draft you approve is exactly the right fit.
Frequently asked questions
Do I need a paid account for a personal AI assistant?
You can set up a Claude Project on a free account too, with limits on the number of projects and usage, current pricing is at Claude Pricing and Plans Explained. Building your own Custom GPT in ChatGPT usually requires a paid account, a free account is typically enough only for using GPTs other people have already published.
Is my data protected with an AI assistant?
That depends on the provider and your own settings, there's no blanket guarantee. Don't upload sensitive or third-party data you don't want stored there, and check before attaching knowledge which data the role actually needs. This isn't legal advice, it's a practical rule from building my own setup.
How do I create an AI assistant in Claude?
You set up a project, write a system prompt with a fixed role and rules in it, and attach your own documents as knowledge. From then on, Claude answers new requests within that role, without you re-explaining the context every time.
How do I create an AI assistant in ChatGPT?
You build your own Custom GPT: a role description in your own words, fixed instructions, and uploaded knowledge files the GPT can cite from. The setup resembles a Claude Project, the exact interface differs.
Can I build my own AI assistant without prior experience?
Yes, all three paths described here are meant for building without coding skills. What you need is a clear idea of the role and a few of your own documents as knowledge, not a technical background.
Where to go from here
Set up a first assistant for the one task where you write the most text and can check a draft the fastest, for many people that's email, as with Sonja. Once you want to understand the difference between an assistant like this and a full agent more precisely, the breakdown is in AI Agent, Chatbot, or Assistant?, and the path to a more independent role with a personnel file is in Creating an AI Employee: How to Start.
The eight ready-made packages, from Sonja's assistant role to the agent roles in the rest of my workforce, are ready to use in my community: Community.
