Taking Notes While FSD Drives: The In-Cabin Problems Drivr Handles

My Tesla does the driving under FSD (Supervised), I do the thinking, and until recently the car's assistant kept getting in the way of writing any of it down.

A lot of my best thinking happens on a drive. FSD (Supervised) handles the road, I keep my eyes on it, and a problem I've been chewing on for days finally starts to make sense.

Then I get out of the car, and most of it is gone.

In-car Grok is good company on a drive. If you haven't seen it, Mike P's clip is a quick look.

But note-taking in a moving car has its own problems, and a general-purpose assistant wasn't built for them. So I built Drivr, a Grok Bot that does one thing: it patiently listens while I think out loud, then files what I said into a memory doc. I use Coda (Superhuman Docs). Here are the cabin problems it's built to address.

But note-taking in a moving car has its own problems, and a general-purpose assistant wasn't built for them. So I built Drivr, a Grok Bot that does one thing: it patiently listens while I think out loud, then files what I said into a memory doc. I use Coda (Superhuman Docs). Here are the cabin problems it's built to address.

Every drive starts from zero

In-car Grok doesn't remember yesterday's drive. Each conversation starts fresh, with no idea which threads are open, what I decided, or what's blocking me. Other Tesla owners have noticed the same thing.

Drivr keeps a durable memory outside the car. Every open thread has a row in a Context table that says where it stands: the goal, why it's urgent, progress, decisions, and what's next. At the start of a capture, Drivr loads those rows before I say a word.

So on the next drive I can say "Hand off to Drivr. Pick up where we left off," or "Let's have a Drivr session." Because I asked, it answers in a sentence or two, built only from what I actually said earlier.

The assistant wants to talk

Conversational assistants answer or interject incessantly. They're programmed to please, and like the LLMs that drive them, they can be distracting. When I'm thinking out loud, a helpful reply to a half-formed thought pulls me off the thread, and I start talking to the assistant instead of working the problem.

Drivr is built to stay out of the conversation. During a capture it says nothing unless I ask. No "got it" after every pause, and no follow-up questions.

Thinking out loud is full of pauses

I don't think in clean sentences. I trail off, go quiet for a mile, say "anyway... so yeah," and pick it back up later.

Drivr treats a pause as thinking. Silence, a fragment, or a half-sentence never ends a capture. Only a cue does: "wrap it up," "that's it," or "file it." If I say something I regret, "scratch that" drops the last fragment.

This is the part I care about most. You need a system that just knows we're in a unique mode. The agent must listen and quietly capture. It should speak only when asked, and keep affirming that it's listening even when you've rambled on while crossing an entire city.

Drivr is built for that mode. If the platform needs a reply mid-capture, the reply is one word: "Okay." During a long ramble, it uses heuristics to reassure me that it's actively listening.

No screen and no keyboard

FSD is supervised, so my eyes stay on the road. Anything that needs me to read or type is a distraction.

Drivr never needs a screen. Everything happens by voice, and when I end a capture, the only thing I hear back is one short sentence with the counts. Something like "Got it: two notes, thirteen to-dos." That was my first real drive, a rambling list about getting our property ready for winter. I checked the rows later, at a desk.

Drivr gets those counts by reading the rows back out of the doc after writing them. If a write fails, it says so in one line and holds the items to retry on the next capture.

Speech-to-text mangles entities and names

Speech recognition in a moving car isn't perfect, and it's worst on project names, people, and places. Even the bot's own name comes through as "driver."

Drivr keeps a Glossary in the doc with each term and the ways it gets misheard. "Driver" is listed as an alias for Drivr, so it gets rewritten before anything is filed. When I fix a name out loud, Drivr corrects the row, adds the misheard form to the Glossary, and answers "Fixed, and noted."

Learning that I can see

A bot that learns while I drive could drift into behavior I never agreed to. Drivr never edits its own instructions. Everything it learns lives in the doc, where I can read it.

Glossary and correction rules update on their own. Each correction also becomes a test ramble. When the same kind of fix shows up across sessions, Drivr writes a proposal, and only mentions it when I ask "anything to improve?" Nothing changes in how it behaves until I approve the proposal and it passes every test ramble.

When the car's assistant cuts in

You reach Drivr by asking in-car Grok to hand off to it by name, or with simple commands that invite Drivr mode, such as "Hey Grok, let's have a Drivr session." Without Drivr, a long pause will allow the car's assistant to jump in. With Drivr, I just start a thought with "note that..." or "capture this...", and Drivr adds it to the capture, stays quiet, and keeps listening until I say "wrap it up."

It works, but the goal is a capture mode I never have to think about.

Try it

I use Superhuman Docs, but you can use Obsidian, Notion, or even just Markdown files. In your first conversation, it walks you through connecting a durable context and setting up the memory doc, including the names it's likely to mishear.

Get the Drivr template.

Get the Drivr template.

If you do your best thinking while FSD drives, try a capture and tell me where it breaks or changes your in-Tesla experience.