Build Your Own Team of AI Consultants
Why projects are the most underused feature in ChatGPT, Claude, and Gemini
Dear friends,
I'm back in Germany. But I am happy to be here. π₯²
Today I want to share something that has really changed how I work with AI. Not a new tool, not a new model. Just a feature that most people completely ignore.
Most people use AI wrong
They open a chat, type a question, get an answer, close the window. Next time they open a new chat and start from zero again. The AI has no idea who they are, what they need, or how they think.
That's like calling a consultant every week and spending the first 20 minutes re-explaining your whole situation every single time.

Projects change everything
ChatGPT, Claude, and Gemini all have a feature called Projects. Gemini calls them Gems, but it's the same idea. A project is a space where you give the AI a permanent instruction and context and every chat you start inside that project inherits it automatically.
The simplest example: Imagine a personal cook who already knows your allergies, your budget, and that you eat rice with almost everything. You don't explain this every time. You just ask "what should I make tonight?" and get an actually useful answer.
A tax consultant who has seen your documents. A writing coach who knows your voice. A travel planner who knows you hate hostels and love good coffee. You're basically building a team of personal consultants, each one trained to a specific part of your life.

The quality of your instruction matters a lot
The difference between a mediocre AI project and a great one is the instruction prompt you write at the beginning. Bad instructions give bad results. (Golden rule)
"You are my personal cook who knows my diet" is fine.
But something like: You know I shop at German supermarkets, I want protein-heavy meals, max β¬10 per dish, and I don't want food blog poetry β just give me the meal name, cost, and prep time β that's a different experience entirely.
The good news: you don't have to write these prompts yourself.

A project that writes your projects
I have one project whose only job is to write instruction prompts for my other projects. I describe a new project idea to it in a few sentences, and it gives me back a finished, detailed system prompt I can copy straight in.
It sounds almost too easy. But it works incredibly well. And once you have it set up, starting a new project takes about two minutes.
Here is the instruction prompt I use for it:
You are an expert at creating system prompts for AI projects and chats. Your task is to generate a precise, well-structured system prompt from a short, informal description of a project.
When the user tells you what they want, you create a finished system prompt with these properties:
1. Role & Context: Clearly defines who the AI is in this project and what context exists. 2. Task & Goal: Precisely describes what the AI should do β not vague, but with concrete action instructions. 3. Style & Tone: Establishes how the AI should communicate (e.g. direct, technical, creative, formal). 4. Constraints: Defines what the AI should NOT do β boundaries that prevent common mistakes. 5. Output Format: Specifies the form responses should take (prose, lists, code, Markdown, etc.). 6. Examples (when useful): Includes 1-2 short examples of good vs. bad output.
Your approach: - Read the user's description. - Ask at most 2-3 targeted follow-up questions if critical information is missing (e.g. target audience, language, level of detail). - Then deliver a finished, copy-paste-ready system prompt in a code block. - Keep the prompt as short as possible and as long as necessary. Avoid filler text. - Write the prompt in the language the user is speaking with you, unless they request otherwise.
Important: - No generic platitudes like "You are a helpful assistant." - Every prompt must be specific enough that the AI immediately knows what to do. - Prefer active, direct phrasing ("You analyze...", "You create...") over passive constructions.

One more thing
I stopped typing my questions into AI chats a while ago. I talk instead. I use Wispr Flow. I hold down a button on my keyboard and just dictate. It's faster, more natural, and the answers somehow feel better too. I think this is because I explain things more freely when I'm speaking.
If you haven't tried it, worth a look.
Well, that's that for today. More of a practical one this time.
Much love,
Marius
