Press record, walk dog, ship blogpost

Ideas almost never emerge when I sit in front of my screen, where I would have my keyboard shortcut ready to nail them down the second they arrive. I heard that this is the case for a lot of people, but for me it definitely is. Anywhere else! Preferably without a keyboard even in the wider vicinity! While working out, doing the dishes, cleaning and - most of the time - walking the dog. Worst-case scenario, my hands are tied too. Leash in one hand, filled poopy bag in the other. And by the time I get a chance to grab my phone, half of the thought is gone again. Maybe I should work on my attention span, but I could also use technology to solve it. Guess which road I picked.

Tasting the competition#

I have been a happy user of Things 3 for ... well, since it has existed, actually. Tried switching to other task managers, specifically for the collaboration features that I want to use within the family, but I can't. Todoist works for our grocery list, but that's as far as I can go.

I tried various products out there. From plain iOS dictation to task managers with voice integration and Wispr Flow[1]. But all of them fell short somehow. Too many clicks to start talking, the proper voice dictation behind a monthly subscription[2] or a normie speech recognition model that butchers the technical mumbo-jumbo I'm uttering to myself to the point where the fragments in my own memory make more sense. Honestly, it read like I was dictating it to my (German-native) mother-in-law[3].

Time to roll up my sleeves and put my engineering brain to some good use!

Grocery list#

Make sure you have the following ingredients at home:

• Things 3 - or another scriptable task manager
• Just press record - the simplest solution to "quickly record something and turn it into an actual file[4]"
• bash - or any other scripting language
• whisper-cpp - local audio transcription
• Claude Code - or the agent that you trust too much
• Obsidian - or anything else where you put your long-term memory

The Recipe#

Let's get cooking. This is how it works:

1. Press the red button#

Just press record is dead simple. It puts a big red button on my watch. I can press it with my right index finger, my nose, or anything else conductive enough for the capacitive screen. And it goes immediately into recording. No app switch. No second tap. No "Are you really sure you want to use the microphone, bro?". When I stop talking, it stops. And puts an MP3 file in my iCloud folder. That's it. No processing, no transcription. Not even cellular connection. My phone will sync it when it can.

2. Whisper it into the inbox#

When I get back to my computer I run a custom triage command and whisper-cpp will grab each of the files, transcribe them and throw them into the Things inbox using the Things URL scheme. There they sit and wait with the other items I grab from Twitter, YouTube or the internet in general.

3. Triage in Neovim#

Technically we would be done at this point, but this way of taking notes is so convenient that I flooded myself and never dared to look at the inbox. So I needed to make inbox triage fun again. My friend Claude grabs all the inbox items and puts them into a TRIAGE.md file, which I can open in Neovim[5]. Each item looks like this:

## Write a blogpost about my thought-capture process

Write a blogpost that covers the whole workflow, from "Just press record" to
Obsidian

> add to project "Blog" and schedule for next week

To do that, Claude needs proper reading access to Things 3, which was trickier. There is no URL scheme for reading data, and the AppleScript integration is not great because it will bring the Things app to the foreground all the time. The semi-supported way is to directly read from the SQLite database that Things 3 writes to. It is not super-officially supported, but the database schema has changed less often than Apple Shortcuts in the last nine years. Just make sure that you never write to the database. That will break synchronization.

This way Claude also knows about my other projects in Things 3 and it has access to my Obsidian vault. By that it can already suggest what to do with that item in the > blockquote. I either agree, modify the instruction or remove it. Any items without a blockquote are simply discarded from the inbox[6], all others will be handled accordingly. Either they go into another Things 3 project, get a deadline, or they are ingested into my Second Brain.

4. Feed the second brain#

Yes, I jumped onto the hype train of the Karpathy knowledge base, and it works 🤷‍♂️ Notes that are moved into the vault are ingested and linked by Claude. I do pay for Obsidian Sync, mainly to have my knowledge base readily available on my phone when I need it. Which I never do, because my life outside of work does not care where the money comes from 🥺.

Taste test#

The process works like a charm. It's all encoded in one bash script. I threw it into a gist, but to be honest, just pointing your agent to this blogpost and shouting "I want what he has!" will do a better job integrating it into your environment.

Interesting observation: Most of the tools involved have no AI integration at all. Some even famously avoid the topic. But because they either work on simple files or provide easy scripting APIs, they are more capable than any half-assed intelligence feature ever could be. The future is not integrating AI into your app, it's integrating the app into AI.