How to Build a Personal AI Assistant That Works with You (Without Programming)

Or: How I went from “New Chat” to “New Task” — from chatting with AI to having it work with me.

ChatGPT, Claude, Gemini, & Co. are designed to keep you in their chat. Convenient for you, profitable for the providers who control every step to their advantage. And it's not a conspiracy, don't worry, but a normal approach aimed at retaining their users, understanding what they're doing with such a new technology, and finding new ways to monetize the colossal investments that AI requires.

But in this seemingly inevitable approach, the question I ask myself more and more often, especially in companies, is "What if the value remained with us?"

Like many others, I started asking myself this question several months ago. And since then I've never stopped creating and redoing environments, setups, and projects to find the ideal solution.

The Context

A few months ago, I needed to experiment with translating and adapting a large amount of material from English to Italian. This was primarily an editorial task, requiring zero computer code. It was a highly complex task that required the efforts of multiple AI agents over a long period of time to be considered effective.

And it couldn't be the classic 'GPT Chat': I couldn't trust the opaque memory of the various AI Assistants, I couldn't start from scratch with a thousand instructions for every new conversation, and I didn't even want to do everything in a system developed using APIs.

The most important word to keep in mind when working with AI is CONTEXT . Context is the information you provide to AI to solve your problems; it's where you explain who you are, what you do, where you're going, and why. And if you're good and lucky, copying and pasting, cross-referencing projects offered in chat apps, supporting mini-agents like OpenAI's GPTs, and a little luck, you can often put together coherent work.

But this greatly distorts the way we are used to working , especially because the various AI system manufacturers are pushing hard on the 'Chat Based' platform in which the system takes care of memorizing the interactions you have had in many conversations (which you will have noticed: there are too many, difficult to find, organize, share) stored somewhere in the cloud.

Every time you download a file, the flow is broken ; generated files and conversations will forever exist in different places, and rejoining them will be a hassle. Essentially, we've been suffering for too many years from poor UI and UX that distract us while forcing us to remember where to click or how to activate thousands of functions.

Think instead about the work you do with your teams.

We take too many things for granted when we work with colleagues: because each of us has a context, a history, a unique knowledge of the company, of our own role, of that of our colleagues, and of "how things are done." Everyone, as Luciano Floridi (a leading philosopher in the digital world) would say, has their own semantic capital that they continually modify, nourish, and compare with others. When you talk to a colleague, your manager, a supplier, or a regular customer, you already know everything that happened before and you work incrementally, producing results.

When you work with the GPT Chat on duty, no. And that creates a lot of problems.

Now, with this work mode, once configured, you'll be able to work as you're used to (even with the presence of AI). Because the AI will live in its context and will be able to fetch the right thing, when needed, on its own.

Imagine a colleague with whom you share a folder. You don't have to explain the context every time. They already know everything. You just have to pick up where you left off yesterday.

Let me explain better.

The intuition

I wanted to attempt an impossible task of translation and recontextualization in Italian for an online course.

Before my eyes was the newly developed Claude Code, an environment designed for programmers, which runs in a terminal. It wasn't decidedly user-friendly. But it gave me the ability to work with several specialized agents, feeding them a huge amount of data. I just had to direct them to write text instead of code.

I fed them the folder on my Mac where the material was and a simple command (more or less): "Translate and adapt into Italian, work in parallel and make sure you do a good job."

The result? After an hour, QuCì, the quality control agent, asked me to validate the result, which was already on my computer, in a folder, so they could proceed with publication. Except that it contained six hundred pages of highly complex content, translated, reviewed, and annotated to explain to me, step by step, how the other agents had worked. Six hundred .

The pages are still there; checking them would have taken too long, and the project is still looking for an author. But the intuition behind it has changed, at least until the next one, my approach to AI .

I was no longer "chatting with AI," but coordinating a production system made up of operational agents . Chats were merely, as in the real world, a tool for achieving a work result that the agents had to take care of. A means of communication, not a way of working.

In the work environment I created in the following weeks and from which I am writing this post, I have now brought a lot of actions: sending or reading emails, writing documents of various types (Documents, Presentations, Spreadsheets), activating browsers and navigating, activating apps on the Mac with MCP, etc.

I started focusing on results and no longer on how I wrote the prompts and one day a few weeks ago, the name 'Outcome Oriented' was born almost by chance for this way of working, Oriented towards Results rather than restricted to a chat.

Let me try to explain myself better.

Chat-Based Paradigm (the GPT Chat Approach)

Your goal is to get better responses in conversations. The value lies in the chat itself. Ask, get, move on to the next question. The conversation is the product and is almost always in the cloud, managed by the vendor in question.

You have to move everything from your PC to the Cloud if you want it to work or use integrations like MCP (IYKYK: I won't talk about it now so as not to further complicate this post)

Outcome Oriented Paradigm (the approach I propose)

Your goal is to work within systems that produce tangible results, with full control of the context and tools available, minimizing the amount of information you pass on that remains 'out of your control'.

The difference is substantial: the value lies in the documents, files, and projects you create. Conversations are instrumental . The concrete output is the product. AI directly accesses your files and executes commands on your hardware to solve the tasks you give it.

To start something new , stop clicking New Chat and ask for a New Task. And that changes everything!

Chat can get lost just as much as the words you share with your collaborators while completing a task. It doesn't matter; what matters is the result.

The difference manifests itself on three conceptual levels:

	Chat-Based	Outcome Oriented
Where does the value end?	In conversations (which you then archive and forget about but are a goldmine for Service Providers)	In folders and files that stay with you, accumulate, reuse, and in which you can find new value.
Your role	"Users" who query a system made by others.	We create what you need to do by leveraging different AIs.
The memory	You hope that AI remembers (opaque memory) by relying on systems you don't know.	You define the rules in folders and files that you control 100%

Here's the crucial difference: while Chat GPT, Gemini, or Claude Web live "in the cloud," Claude Code works on your computer. On your documents. With your rules. Keeping in mind that some care will be needed (again, If You Know, You Know: there are huge cybersecurity issues that I'll discuss later).

	Chat-Based	Outcome Oriented
Where the tools 'run'	On the cloud	On your Mac or PC
Who should take care of backups?	The supplier	You
The control	No Brain	Attention is needed

Let's get practical

"Claude Code is an agentic coding tool that lives in your terminal, understands your code, and helps you program better and faster ."

It was created for developers, lives in the terminal command line (the old "C:\>" prompt for the less experienced), and is intended for those who write code.

If we change the sentence above to

"Claude Code is an agentic coding tool that lives in your terminal, understands your documents and instructions, navigates the complexity of your computer and its folders, tools, and applications, and helps you work better and faster."

we get a tool that we can also use for everyday activities

What this essentially means:

• Reads and writes files on your computer (reports, presentations, documents).

• Work on your computer by directly running applications and code.

• Remember who you are and how you work (permanent instructions, 100% managed by you).

• Respect your organization (separate folders and projects, not a single chat).

• Work with sensitive files without uploading them to the internet where you'll forget them.

• He learns through experience based on what you tell him to do or not to do.

• And much more that I won't explain now (I'm lazy and it would take too long).

He's just like the colleague you spoke to yesterday. He already remembers everything.

The Limits of "Memory" Assistants

Some might think that ChatGPT has memory, Claude has projects, Gemini remembers past conversations! And the cloud is easier to manage, especially with MCP.

True. Nowadays, memory is something everyone provides (I wrote about it here ), but there's a fundamental problem: you don't control it .

You never know for sure:

• What they really memorized and what they didn't

• If that conversation from three weeks ago is still in his "memory"

• Because today you seem to have forgotten your tone of voice that "reminded" you of yesterday

It's an opaque memory. You can't inspect it, modify it, or structure it to suit your needs. It's like having a colleague who "maybe" remembers things, but you're never sure what. He's distracted by a thousand things, and his priorities are set outside the work environment you're interested in.

Think about concrete scenarios:

• Ghost information : The AI cites a piece of information it "remembers" from a past conversation. Unfortunately, that information was already incorrect at the time, and now it's repeating it as if it were true.

• Project Contamination : You've been discussing two different clients in separate chats. The AI is mixing details from one with the other. You don't notice.

• Unfounded Confidence : He confidently answers you using "what he knows about you." But you can't verify what he knows, where he got it from, or whether it's up to date .

• The silent bias : He "learned" that you prefer a certain style. But he learned it from a conversation where you were joking. Now he applies it all the time.

The problem isn't that he forgets. It's that we don't know what he remembers, because, if memory returns at the right time ... And when he's wrong, he does so with the same certainty as when he's right. You find yourself hoping he remembers, instead of knowing for sure what he knows.

In the first case, the almost certain result is a "SLOP" output, meaning carelessly assembled, often randomly assembled pieces of content. This issue explodes in multi-agent systems that can generate content in 10% of the time, and then require 200% of your editing time.

Support agents

The latest AI platforms are able to offer increasingly comprehensive tools that not only respond but also perform actions. And we're calling them AI Agents.

What is an AI agent? The shortest definition I can give is: a Reasoning Language Model (RLM) with access to digital tools that works in a loop until it completes the task you give it.

I've talked about it at length here and here . But you can find a lot of literature online.

In an Outcome Oriented environment, you'll need AI agents to operate quickly. And they operate with simple commands that underpin many of the activities you perform in the digital world:

• Search for text and documents (grep, find, mdfind)

• Read and write text in files (cat, less, nano, echo >)

• Run commands on the operating system (bash, any command)

• Perform web searches and read the contents of found files

• Manage to-do lists (Task list or To-do)

• Connect to any other instrument, local or web, via MCP.

One of the wonderful things about this environment is that you can have AI create agents specialized in certain roles (Data Analyst, Presentation Creator, Proofreader, Professional Chart Creator, Email Responder, Sales Proposal Writer) simply by writing their job descriptions.

You can create specific skills to perform tasks (such as taking meeting notes, adapting your presentations to your brand identity, translating documents following certain rules, accessing your ERP or CRM to download or upload data, or executing small pieces of code for repetitive tasks.)

All this in an environment that will never give you (big) surprises.

So what?

Want to know how? There's just one problem: this post is actually the longest preface I've ever written!

To give you the opportunity to work in an Outcome Oriented manner, there's a lot to talk about and get your hands dirty, and especially at the beginning, it will take some effort. Figuring out how to move from the "hope" of achieving good results on a commercial AI assistant to the confidence of producing valuable results in your own customized work environment requires some commitment.

But the results are interesting and I could tell you how:

• Build the system step by step, even without being a developer

• Equip him with skills — like in the Matrix: "I know kung fu" (Excel, PowerPoint, PDF...)

• Make it multi-agent — not an assistant, but a digital team that coordinates

• Connect it to your systems — Notion, Email, ERP, CRM, OneDrive, other Agents.

• Create complex workflows — Elaborate workflows that mix AI agent capabilities with code execution (strictly AI-generated)

Do you like the idea?

But I need your help to decide how to proceed:

• A book ? (paperback or ebook, complete structure, permanent reference)

• A series of articles ? (at least 5 parts, feedback on progress)

• An online course ? (video + exercises + community)

• A prompt ? (So Claude can explain to you directly how to do it - best wishes! -)

• Dedicated workshops ? (corporate training, hands-on, online?)

Write to me in the comments or contact me directly on the site. In these times of excessive AI-generated content, it makes sense to think twice.

What intrigues you most? What scares you? What would you like to understand before anything else?

We'll build this guide together.

Max

Enjoy AI Responsibly!

(P.S. this post has been auto-translated by the AI environment... I've presented in this post)