I’ve been using Evernote since 2011. Fifteen years of notes, and somewhere in there I built an organizational system that’s basically Getting Things Done, the Secret Weapon method, and the Eisenhower matrix all smashed together and run through a blender that only I know how to operate. It works. It has never once made sense to another human being, and it doesn’t need to. Oh, and hell yes, I used AI to help write this, but it’s still my voice, as you’ll see below.
The problem was never organizing. I’m good at organizing once the thought is captured. The problem is that my brain produces thoughts faster than my hands can type them, and every time I stop to write something down properly, I lose two more thoughts behind it. So the actual bottleneck was never “where does this go.” It was “how do I get this out of my head without editing it into something that isn’t what I meant.”
Here’s the pipeline I built to fix that, and why voice turned out to be the missing piece the whole time.
Voice beats text, and it’s not close
I use Whisper for dictation on my phone and computer. Still my words, still my voice, it doesn’t rewrite anything unless I tell it to. That alone was a big upgrade over typing. But there’s a second, less obvious reason voice works better for me than text, and it’s not about speed.
When I type instructions to an AI, or even to a coworker, I have time to smooth it out. I reread it. I soften the parts that sound like I’m annoyed. I turn “just do the thing, why is this hard” into something diplomatically padded and, honestly, less clear. Voice doesn’t give me that chance. What comes out of my mouth is what I actually meant, curse words and all, before my brain gets a chance to sand off the parts that make me sound like a guy who knows exactly what he wants and isn’t going to pretend otherwise. I’m not pompous in my own head. Typing sometimes makes me sound like I’m trying to be. Talking never does.
That’s the whole reason a second voice tool made sense on top of Whisper.
Enter the ring
Whisper still needs me at a device. Phone in hand, computer in front of me. It doesn’t help when I’m walking in the morning, driving, or standing in the garage thinking about a client deadline while I’m doing something with my hands. That’s the gap the Vocci ring fills. I tap it, I talk, it’s gone. No phone out, no app open, no “let me just type this real quick” that turns into me not typing it at all because I got distracted by literally anything else.
The ring records the thought. What happens to it after that is the actual system.
The harness
I run Claude as a harness: the thing that directs other robots rather than doing every task itself. Underneath it I’ve got Codex and a local Ollama instance, and Claude manages the handoffs between them through a self-directed Kanban board. I dispatch work to the harness two ways: through Cowork when I’m at a computer, or straight from my phone when I’m not. I could wire that up as texting if I wanted to get fancier about it, and probably will eventually.
The point of the harness isn’t that I have a lot of robots. It’s that I only have to talk to one of them. I don’t route tasks to Codex or Ollama myself. I tell Claude what I want, and Claude figures out who’s best suited to actually do it. That’s the same principle as the ring, just one layer up: reduce the number of places I have to translate my own intent into something a machine understands.
What actually happens every two hours
There’s a cloud-based trigger that wakes Claude up every couple of hours and has it pull whatever I’ve recorded on Vocci since the last pass. This part runs regardless of whether my laptop is on, which matters. I didn’t want a pipeline that only works if a specific machine happens to be awake when I am.
Here’s the part I think is the actual trick: Claude doesn’t just trust Vocci’s own AI-generated summary of what I said. It goes back and rereads the raw transcript itself. Then it goes and looks at my existing Evernote notebooks and tags (not a hardcoded list I handed it once, the actual current state of my taxonomy) and decides where the thought belongs based on how I’ve already been filing things. It’s pattern-matching against fifteen years of my own behavior instead of applying some generic productivity framework I’ve never agreed to.
From there it creates a new note for that voice capture, in the notebook it decided fits, with the tags that fit. If I gave myself an explicit instruction, something like “remind me to follow up with so-and-so,” that becomes a task with a due date, defaulting to same-day unless I said otherwise. If I didn’t give myself an explicit task but the content clearly implies one, based on how I’ve handled similar notes before, it’ll add that task too. And if it genuinely can’t tell whether something needs action, it doesn’t guess. It drops a simple task to review the note myself. Nothing just sits there silently filed and forgotten. Worst case, I get a two-second task telling me to go look at something.
There’s also a dedup layer running underneath all of this: an append-only log of every voice capture the pipeline has already processed, checked against the note itself in case I’ve since renamed, merged, or deleted it. I did not think about this until the pipeline actually double-filed something once (yea, I deleted the same note like 4 times, haha). Now it’s belt-and-suspenders: two separate ways to catch “I already handled this,” so a recording never gets filed twice even after I’ve gone in and cleaned things up.
What this actually looks like
A few real (details changed) examples from the last couple days:
I was thinking about swag and dinner logistics ahead of a trade show and just talked it out while I was doing something else entirely. It came back as a note in my events notebook, tagged the way I tag every other show-prep note, with a task to follow up on vendor pricing.
I mentioned, almost as an aside, that I needed a creative brief for a client’s marketing initiative using an existing location’s signage as a case study. It filed itself under client work with the right tag, and flagged a task to actually go write the brief, because I’d stated an intention, not just a thought.
I was thinking of a recipe idea for something to cook this weekend, completely unrelated to work. It landed in my personal/cooking notebook, no task attached, because there was nothing to act on. Just something to keep.
Even the idea for this post came out of a voice memo while I was walking around, complaining about people who gatekeep AI or oversell $500,000-a-year multi-agent setups that, from what I can tell, mostly do nothing. That memo got filed, turned into a task with a deadline, and here we are.
Evernote still does its own thing on top
Once something’s filed, I’m not done leaning on AI. Evernote has its own AI baked in now, and I use it to further clarify notes or smash near-duplicate ones together: a second pass, inside the tool itself, on top of whatever Claude already did on the way in. It’s not redundant. It’s just another checkpoint.
The actual point of all this
None of this is about how many robots I’ve got running. It’s about never having to be the bottleneck on my own thoughts, while still being the one who catches anything that got filed wrong. Every note the pipeline creates, every task it assigns me, I can go look at. I can see exactly what the robots did and override any of it in about ten seconds. That’s the part that actually matters: not the automation, the transparency underneath it.
I know a lot of people talking about AI right now are trying to sell the idea that you need thirty-five subagents and a six-figure compute bill to get anything out of this stuff. I don’t buy it, and honestly, I don’t think most of them are actually doing that much with it either. What worked for me was looking at what I already do (talk instead of type, file things a certain way, catch myself mid-thought while driving) and building the AI around that instead of rebuilding my life around the AI. Know your own system before you hand it to a robot. Then let the robot carry the parts that were never the interesting part of the work anyway.