Voice notes

How do you pour out the knowledge sitting inside you? On how LLMs help us write down our ideas.

The problem

Many of us probably struggle with a flood of ideas that we jot down on paper and then lose. Or we already have a full sketch of an idea in our head, but no time to flesh it out or carry it through among all our duties.

I had a few unfinished topics:

My website, which I knew how I wanted to change, but I had no time, because it was a purely creative project with no profit potential.
Becoming more active in the industry and sharing with a wider audience the knowledge I share with friends and people around me. Because I love sharing knowledge, and I think “know-how” is worth sharing, since it is not a big competitive advantage. For me, the advantage is knowing how to use it.
Growing my company’s marketing.

Today I want to tackle topic no. 2. In 2022 I started writing blog articles. Plenty of knowledge. I even had everything prepared and uploaded to the site, but it was missing the final polish. Filling in the substance. Fixing mistakes. Translating into two languages. And so the topic got stuck. Because there was always something more pressing.

Today, on the day of writing this article, I have already managed to write three. And I am writing this one purely for the joy of writing. I used AI for it, but not the way you think. Every sentence is still my own.

The idea

The idea for this process came after I read a post by Ilya Sutskever saying he had started dictating tasks by voice to his agents/LLMs/AI. He described using simple speech-to-text to turn speech into text. And the most important thing about it was that he did not worry about formatting. He simply poured out all his thoughts, even broken sentences. He poured out his whole process, knowing the LLM would sort it out and understand it anyway. That is when it clicked for me.

The solution is TTS + LLM

Let me sketch my current process. While driving the kids or coming back from work, I start recording and make voice notes on a specific topic. I drop the note into a speech-to-text tool such as Orukeet or Parakeet. They work great with Polish. What’s more, I use the simplest models. They spit out a transcript – with errors, with jumbled words. It is my thoughts poured out without any filter. Then I bring in the LLM. I have a prompt that tells it to stick to my text, but to:

fix style and punctuation,
fix unclear words that should be different (given the context),
merge places where I covered the same point twice into one,
split the text into coherent paragraphs and, if needed, regroup sentences that belong in another paragraph (it is my stream of consciousness; sometimes I jump between topics or come back to an earlier one).

And that is how I get a nicely composed blog post. I embellish it with images and charts myself.

But that’s not all…

Google had just made skills available, so I created a skill that adds metadata such as the title, tags and “summary” to the article, in the right format for my blog.

But that’s not all!!!

I already had a free app that used Orukeet to create notes, but I still had to move them over to Gemini… So I asked an agent to build me a transcription app with Orukeet that formats the result straight away with a given prompt. The agent delivered it in 5 minutes. A ready app in Rust and Tauri, using Orukeet and gemma4:e4b, a small model from Google that runs entirely locally. It works perfectly.

I won’t even share it here, because I think each of you can already build it on your own. I use it every day now.