How I Built My Second Brain In 3 Hours
And use it to turn thoughts into actions in mere seconds
This week I built a personalised second brain.
The starting point was that I wanted to do more with voice—as of last week I was still typing notes like a troglodyte.
The average person types 40 words per minute but speaks at 150 WpM, so voice allows you to pass much more context to your AI in the same amount of time. And the more nuance you add to your AI instructions, the better the quality of the AI-generated output will be.
The idea capture system I need had to capture four types of records:
Observations: inspired by Matthew Dicks’ “homework for life”, I record things that happen in my day-to-day that I want to remember over time.
Business ideas: ideas on how to grow and improve my online businesses.
Content ideas: seeds of thought for different types of content (Substack notes, newsletter posts, Youtube videos, shorts and Linkedin posts).
Readings: comments and ideas gathered from reading books or watching video content. These needs to be tagged by origin so I can consult the source material.
This is how I captured ideas before:
Open Notion app, browse to page (business, content, observations, readings).
Create a new Notion database entry manually and write down the idea.
This was a cumbersome and involved process.
For example, my observations Notion database has hundreds and hundreds of entries and scrolling down to the last row to append a new entry was a bit of a chore.
The first thing I tried was to hook up my AI assistant (Google Gemini) to Make.com.
I liked the idea of not having to install another app on my phone.
Turns out AI assistants are bad at recording things.
But because AI assistants are designed to be conversation partners, they have a strong tendency to talk back when I just want them to record things.
Both Gemini and ChatGPT offered plenty of unwanted advice on the first few tries, even when I explicitly asked them to only transcribe what I said.
After some digging I was able to leverage the native Google Gemini—Google Keep integration. By asking Gemini to create a Keep note it would just transcribe the literal content and automatically create a Keep note in my Google workspace.
The workflow was still pretty clunky though: I would need to explicitly tell my Gemini AI assistant to “create a Keep note” each time I wanted to record an idea.
And because the Google Keep API doesn’t have a good integration with Make.com this setup turned out to be a dead end. Google Keep would need to be polled from Make with the Make.com Google Keep module. This introduced both a 15m delay in processing new notes and additional costs from running the Make.com scenario over and over again to poll (check for) for new notes.
It took me about 1.5 hours to figure all this out—let’s call it the research phase—which is about half of the time I spent putting together my second brain.
After doing more research, I settled on a dedicated recoding app instead.
There are several good options out there, including Otter.ai and Notis.ai.
I ended up choosing voicenotes.com because of the quality of the transcriptions and the ease with which you can set up integrations with other productivity apps:
Voicenotes.com does have a couple of drawbacks that you need to be aware of:
They use OpenAI and Anthropic AI models to process voice data, so your voice recordings will be sent to these third party AI model providers. This means that for classified business data or strictly personal content it might not be the right choice. Since most of the notes I take will end up in the public domain one way or another I didn’t care—but it’s good to be aware.
There are some user-reported issues with data loss for longer recordings. Again, not that relevant for my use case, since I only plan to record short notes.
The Voicenotes free tier limits recording time to one minute. I ended up paying for a monthly subscription to not run into this limit. Their monthly plan costs EUR 15/mo. I’ll probably switch the annual plan (EUR 90/y, i.e. EUR 7.5/mo) if I’m happy with my setup.
To be clear: you can build the solution below using just the Voicenotes.com, Make.com and Notion free plans. The only thing you would need to pay for are API credits for the Google Gemini and Claude Opus AI models used in the Make.com scenario below—depending on your note-taking volume and how you set up your own second brain these API credits will probably cost you between a couple of cents to a couple of euros per month.
Building my second brain
The idea capture automation workflow I ended up building looks like this:
The Voicenotes app on my phones records notes and transcribes their contents.
The creation of a note in Voicenotes triggers the following make.com scenario:
Let’s walk through each of the elements in the Make.com scenario above:
The Webhook listens to my personal Voicenotes.com workspace and will trigger the scenario when a new note is created or an existing note is updated.
The Classify note Gemini 2.5 Flash model completion figures out what type of note this is: an observation, a business idea, a reading, or a content idea:
Here is the prompt I used in the Classify note Gemini 2.5 Flash completion:
Based on the transcript below, if it is a content idea output just the word "content". If it is a business idea, output just the word "business". If it is an observation about life in general, output just the word "observation". And if it is commentary on sources - for example from reading textbooks or watching content - output just the word "reading".
The user might present hints in the transcript how it needs to be labeled (for example, readings will always contain the text "reading <title xyz>").
This is the text of the transcript:
{{5.data.transcript}}
The Router has branches that route the transcription of the incoming voice recording to the right arm using simple conditional logic:
The four Create a Database Item Notion Make.com modules create new database records in Notion databases (observations, content, business ideas and readings).
Gemini 2.5 Pro completion modules like Strategy analysis or Source determination enrich the voice note transcription with additional AI analyses. These are either written to Notion or used in follow-up automations.
For the “content idea” branch, I added a workflow that automatically generates first drafts of LinkedIn posts from any content idea I record with Voicenotes:
The entire scenario was very straightforward to set up—it took me a little over 60m.
The only issue I ran into was that new Notion database entries have a maximum size of 2000 characters and the Strategy analysis module in the “business ideas” branch returned output that was much greater in size.
This is the prompt I used in the Strategy analysis module:
Your are a business strategy expert. You are helping Jonas strategize. Jonas is a Dutch AI marketing automation agency owner, content creator and freelance ML engineer. Jonas has been an ML engineer for the past 10 years, the last 5 of which as a freelancer. He holds a Masters in Marketing Intelligence and a Bachelor in Philosophy of Language, and has done a lot of client work on marketing technology, natural language processing, recommender systems and generative AI projects. He write The Circuit Substack weekly newsletter on AI and automation, and has a Youtube channel called AI in practice that helps people get more out of AI using real-world insights.
Evaluate the business strategy idea outlined below:
{{5.data.transcript}}
Find the weak points and strengths of the idea, and figure out the best way to execute on the idea.
The date today is {{now}}
Write out the report as a SWOT analysis + a plan of action with the best chance of success.
When asked, Gemini 2.5 Pro came up with a nice solution to the 2000 character limit for Notion database entries: leverage the fact that each Notion database entry is also a Notion page, and write the content to the newly created database item page instead.
This worked beautifully, as demonstrated below:
Of course the follow-up actions to the voice notes in the second brain scenario above are just some examples of what you can do with this setup.
Given that make.com has integrations with pretty much any business system on earth, you can use the same setup create a second brain that updates CRM records, drafts emails, schedules meetings, does research, or even automatically creates entire apps using AI coding tools like Lovable.
A word of warning though: even though it is technically possible, I’m not using AI to directly interact with business systems. Each branch of this automation ends with a human reviewer—me—deciding on what the best next step is.
The main difference with my previous setup is that I can now save ideas in seconds rather than minutes, and that I can automate most of the follow-up actions so that when I do have time to work on ideas, I can start immediately.
Last week in AI
The White House released America's AI Action Plan, a national strategy to guide AI development and deployment. The plan focuses on promoting innovation while managing risks in critical sectors.
Swedish AI vibe coding startup Lovable achieved $100 million in annual recurring revenue just eight months after launch number four.
An advanced version of Google's Gemini Deep Think AI system made history by achieving a gold medal standard at the International Mathematical Olympiad, earning 35 out of 42 possible points. The AI successfully solved complex problems at a level matching the world's most brilliant young mathematicians.
Alibaba launched Qwen3-Coder, a high-performance open-source AI model designed specifically for coding. It provides a strong, freely available alternative to proprietary AI models and AI coding tools.
The online coding platform Replit introduced Queue, a feature for managing complex tasks with AI agents. It allows users to line up a series of jobs for an agent to execute sequentially, even while the user is offline.
GitHub launched GitHub Spark in public preview for Copilot Pro users. Github’s entry into the vibe coding space lets you move beyond simple code suggestions to brainstorm ideas, plan project architecture, and build applications from scratch.
this is cool...thanks for the detailed guide, man
i tried manus and crew but i think i probably prefer python and bash.