When my job ended in April 2026, I set aside dedicated time to build and learn with AI. This page shows where that got to: the system I run my work on now, how it grew, and what I learned.
10 Nov 2026
WorkLifeSystem backboneAnthropic's sideRetiredCircle size: dated changes. Hover or tap a circle for its story.
Each item appears on the date it was added. Items added in the last three weeks before the slider's date are outlined.
Voice conversations a month
* November runs to the 10th.
What my voice writes were about
Writes that voice sent to the shared record each week, from 12 October, when voice could first write to it. "About the system" is a rough filter on where each write went, not checked by hand.
A few things I learned along the way
Build your own test interfacesCustom-built interfaces are great for evals: fully control the inputs and outputs, chain them together, feed in production data. I built them for all my quality work at Waymark, so it was good to hear the evals course teach the same approach.
Good = no more badGet to good on subjective problems like creative quality or chat responses by identifying failure modes, grouping them and eliminating them one at a time.
When expected fixes fail, get creativeI couldn't get multilingual voices to sound right until I told them to speak like someone on the radio in a region-representative city.
Use AI to improve AIWhile building internal test tools in Bolt, I had AI analyze one of my early build sessions: roughly 40 to 50% of iteration time was going to the same recurring failures. From then on, every spec had three parts: core fixes that headed off those repeated problems, components already proven to work (a shared sign-in, model selector and prompt version manager, used where a tool needed them), and additions specific to that tool. Each tool after that launched faster than the last.
Test which model to use where, and when not to use oneWorking on an AI video pipeline, a speed benchmark of seven models from three companies found a different winner for each step. Some steps worked better with no model: plain code was faster and more reliable.
Use AI to remove bottlenecksWhen I took on our AI video pipeline, the codebase was undocumented, the engineers on my team were new to the company too, and the ones who knew the code were busy with other work. The team was unblocked when I reverse-documented the codebase with GitHub Copilot and proposed my own pipeline redesign.
The system now
What I built is a multi-agent system I run my work on. Desktop, phone and voice all read and write one database, including the instructions each session starts with, and it reaches 19 tools and data sources, including email, messages, GitHub, LinkedIn and the web. On a typical day in October I had about eight sessions open at once, handing work to helper agents close to 900 times that month, and some sessions now start and manage others. I can work with it by voice for an hour or more without it losing the thread, and most of what I change by voice now is the system itself.
More reach needed guards. My messages are read-only through a connector I built and published, a hook blocks sending, and changes that don't match exactly wait for me.
Reach and ways of working
I started where most people do: one chat at a time. In May I tried Claude Code, which works on files and remembers what it did, and over the summer my work moved there; by September I was running more Code sessions than chats. Voice began as a side door, for interview practice.
The move from chat to Code was about doing more: more work in parallel, more tools, more reach. AI also unlocks voice as a new kind of interface. Anything that doesn't need enough context to require a screen can now happen much faster. I'm striving for voice parity wherever possible.
Four problems I took on in voice
Most of the pieces are common now: voice assistants that reach email and calendars, memory, dictation into notes. What I worked on were the harder parts.
Long talksVoice conversations tend to lose context as they run long. I can talk for an hour or more across several topics without starting over, and when I checked my long conversations, the replies didn't get worse further in.
My turnVoice assistants have to guess when you've finished talking, and they often cut into a thinking pause. In my interview drills the assistant stays silent while I compose an answer and gives feedback only when I say I'm done.
One recordVoice writes to the same records my desktop and coding agents use, through a check that holds anything ambiguous for me.
BuildingIn one week in early November, most of what I changed by voice was the system itself.