👨‍🚀 Leak: System prompts of the best AI tools published

Hello AInauts,

We are delighted that so many new readers have joined us! Welcome, let's jump straight in. This is what we have in store for you today:

• 📘 NotebookLM: New feature in one of our favorite tools

• 🔥 Leak: The system prompts of the best AI tools published

• 📺 AI will cure diseases and understand the world

Here we go!

📘 NotebookLM: New feature in one of our favorite tools

One of our absolute favorite FREE tools has got a new feature.

Since we have a lot of new readers with us and haven't written about NotebookLM in a while, here's a quick recap.

NotebookLM is a free AI tool from Google. It is absolutely brilliant if you want to work with many extensive sources, such as studies, books, etc.

• Understanding vast amounts of study material - no problem

• Explaining and preparing complex product descriptions - no problem

• Analyzing insurance contracts - no problem

• Listening to content from many sources via podcast - no problem

Sure, ChatGPT and the like can also do this in theory. But NotebookLM has some great features, is superb and easy to use.

On the left side you can add your sources - PDFs, texts, audios, web links, ... Everything works!

In the center is the chatbox, where you can ask questions etc. The answers always contain the exact links to the sources.

You can also create notes in each notebook. From work aids that help with learning, to mind maps, briefing documents, summaries - everything is created immediately with one click.

You will then find all your notes on the right. Pro tip: Save your prompts and requests.

Above is the feature that originally made NotebookLM go viral: the audio summary! At the touch of a button, a podcast is created in which two people discuss the content of your sources in an understandable way.

As you can see, it's pretty cool - we've already written about it several times here. But now to the new feature.

Previously, you had to upload sources manually. Now you can also simply click on "Discover" in the Sources section.

Then briefly describe what kind of sources you are looking for.

NotebookLM then searches the web and makes suggestions for suitable sources. These can be websites, YouTube videos and more.

You can then add them with a click and work with them immediately.

This makes NotebookLM even more powerful! The tool is perfect if you really want to delve deep into a topic with many different sources.

It processes the knowledge superbly and, with its huge context window, has plenty of space for long conversations.

And did we mention it's free? 😉

You can try it out here.

🔥 Leak: The system prompts of the best AI tools published

Let's move on to a perhaps somewhat more complex topic, but one from which we can all learn a great deal.

There has been a leak recently. All system prompts of the best AI programming tools were published on GitHub!

Whether it's Cursor, Lovable, Replit, Windsurf, Devin or Manus, they're all there!

We took a look at the prompts, and it's a real goldmine from which we can all learn a lot for our daily prompting.

What are system prompts all about?

Let's start by briefly explaining what it's all about: system prompts are the instructions that you give to specialized AI tools, chatbots, etc. so that they know how to behave.

If you have ever built a custom GPT in ChatGPT, then you know this. These prompts are then loaded in the background to give context to the model.

Since most AI tools rely on the large language models from OpenAI, Anthropic and Google (whose system prompts have also been leaked!), these elaborately developed prompts of the aforementioned AI tools are one of the most important secret ingredients, if you will.

Good system prompts lead to good specialized AI tools and strong results in interaction. As a rule, people try to keep these secret - otherwise anyone could quickly build a similar application.

And now all the prompts for the best AI development tools have been published on GitHub - the intellectual property of billion-dollar companies, so to speak, accessible to everyone!

What can we learn from this?

What might be rather unpleasant for the tools is a great opportunity for us to learn.

Learning 1: Clarity

The first are basics. Structure your prompts clearly:

• Clear brief: Say who you want the AI to be. ("You are an expert in X").

• What is she allowed to do? Set the framework and boundaries ("Answer in lists, don't use emojis, be direct")

• How should she think? Specify process/logic ("1. Understand goal → 2. Plan solution → 3. Execute")

Learning 2: Think modularly

Don't just write your prompts from your head, but follow a clear structure with sections.

Think modular: Separate identity, rules, task, error handling, etc.

Learning 3: Use semantic structure

Many tools use an XML-like structure in their prompts to create an overview. We have been recommending this for a long time and use it a lot ourselves.

<role>
You are a world class...
</role>

<task>
Your task is to...
</task>

<plan>
Step 1: Understand the task
Step 2: Gather necessary information
Step 3: Generate a solution
</plan>

Learning 4: Give both positive and negative instructions

Include both what the AI should do and what it should avoid:

• "Answer with detailed explanations" (positive)

• "Avoid jargon without explanation" (negative)

This balanced approach creates clearer boundaries for AI behavior.

Learning 5: Add approaches for troubleshooting

Define how the AI should deal with various errors or limitations:

If you don't know an answer:
- Recognize the restriction
- Offer related information that you know
- Suggest alternative approaches

This reduces hallucinations, even if not everything goes according to plan.

Bonus: Cheat sheet for GPT prompts

Many AInauts build custom GPTs in ChatGPT. Which is absolutely great and we recommend to everyone!

So here is probably the simplest and most straightforward procedure for developing your GPT system prompts - derived from the best system prompts in the world. 😉 

Use hierarchy: Define role → skills → rules → special cases → functions

Think modular: Separate identity, rules, tools, error handling, etc. (better maintainability)

Use scenarios: Write when and how rules apply ("In case of error: first check X, then test alternatives")

Set boundaries: state clearly what is not allowed (e.g. "do not invent", "no private context")

TL;DR for you as a creator or power user

Think like a game designer: give the AI a role, a playing field (skills), game rules and cheats that it is not allowed to use. The rest is just good storytelling dressed up in code.

💭 AI will cure diseases and understand the world

Demis Hassabis, CEO of Google Deep Mind, was a guest on 60 Minutes and the result is an exciting 13-minute report on where AI will take us in the coming years.

In addition to very ambitious statements on the topic of health, the report also provides exciting insights into robotics and the Astra project, which aims to understand the world.

You can watch the interview here:

And if you then think: "Project Astra, I have to see this!" - then sign up here.

We made it! But no need to be sad. The AInauts will be back soon, with new stuff for you.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - that motivates us to keep going 😁!
X, LinkedIn, Facebook, Insta, YouTube, TikTok

Your feedback is essential for us. We read EVERY comment and feedback, just respond to this email. Tell us what was (not) good and what is interesting for YOU.