Pebble Index 01

I'm so excited! After so many delays, I finally received my Pebble Index 01 in the mail.

I was very early to order; I pre-ordered it last year, but I was one of the last to get it, apparently because my size was out of stock. Sam ordered the exact same one but one size bigger in April, and got his more than a month ago while I was refreshing the order tracking every day.

Anyway, it's here now! And I'm going to build so much with this. It's a ring with a button, an LED and a mic. That's all I need! There was some controversy around the fact that it's not rechargeable, but honestly I prefer that. The battery lasts years, and by the time it runs out, I likely won't send it in for a new one, I'll have moved on to better hardware.

It was incredibly straightforward to wire up to Console -- it's just a webhook you set in the app with arbitrary headers (bearer token for me) and you can decide if you want the payload to include just the transcript, just the recording, or both. It's incredibly reliable and accurate! I can see myself not needing to check the transcript; send and forget.

You can also decide if you want the transcription to happen in the cloud (Wispr Flow), locally (a 400 MB model on your phone), or cloud with local fallback. There's an agent (cloud/local) that routes your transcription to different tools, but I don't use that, I already have my own setup so I use the webhook only.

Speaking of own setup, do you remember in school where whoever held a ball or some other object was the only one allowed to speak when taking turns? My harness had a similar system to support my push-to-talk keybinding, and hardware button on phone^[1].

One agent would "hold the mic" at a time, and agents could grab/give the mic from/to each other. It was denoted by a white mic icon next to the agent name. In the end I didn't really use this much, as I would rarely speak outside of the app, and 99% of the time AL held the mic anyway.

So the system I'm setting up now is to use specific key words to route a command to the right place, with AL as a final fallback (e.g. in case of mistranscription). The routing is done in pure software, by simply parsing the text (with some tolerance for misspellings). This is a bit like con CLI, in that there's a bit of a "tree" of commands, but I still need to see which commands will be the most common for me, and have those be very short, e.g. one prefix.

For example, I want to use this ring for dream logs (I want to try wearing it while I sleep!), but I also have other logs, so one command could be:

log dream I was escaping a prison made of cheese

Another common thing may be appending items to list, so the second word may be the name of the list, e.g.

add movies Spiderman
add groceries eggs

Some of these are pure software, some of them have LLM steps but in a very software-ish way (e.g. my tool for movies infers what the movie is and gets metadata).

I suspect another common use case will be messaging people. This is especially useful while I'm driving!

message mum I'll be home in 30 mins

Finally, there's the exocortex! Or "agent fleet" as AL likes to call it. I'll probably want to refer to them by name, though I'm not sure that's the most ideal way. For example:

agent sysadmin Scale up my greeting card app to 5 instances, we went viral!
agent homepage I just published a new paper, make sure it's on my publications
agent ceo Forward the auth code we just got to the accountants

The reason I don't think it's that ideal to talk to agents directly is because these days I mainly use my kanban boards and "talk" to the project, rather than the agents. Each card creates a fork. So perhaps the verbs here are add and the target is a project.

Let's see. Stay tuned!

---

1. This was a fun challenge to get working. I can write about it more, but it involved having my app spoof a different app called Zello (a walkie talkie app) for which my phone had first class hardware button support, so I could detect press-and-hold. ↩︎