9/14 - System Snapshot: How I'm Running My Life with AI

I'm sharing this publicly as a working note for myself about the system I'm building. It records what I'm using, how the different activities connect, and the unfinished parts I need to return to.

I'm currently unemployed, living in San Francisco, and trying to keep the things I care about moving while I look for work. It's been a difficult period. Having more time hasn't automatically made it easy to stay engaged with everything I want to do.

One connected system: external reality exchanges experience and attempts with businesses, job search, education and daily life. The businesses develop Busibody. My direction and retained context support the activities; their observations inform Idyllic and the writing queue. Public writing returns to the outside world. — My activities share context and feed the infrastructure and writing around them. I still manage many of these connections.

Job search and Idyllic get most of my attention. The businesses and writing give me work to do with agents, while languages, study, exercise and life with other people have their own reasons to be here. I still initiate much of the work and carry context between these activities.

My current AI stack

I record thoughts on my Plauds while walking and talking. The transcripts preserve how I arrived at an idea, including the questions and qualifications that disappear in a short summary. Alongside years of journals, they supply material for writing and for agents working with me.

ai-organs is the heavy context center. It holds daily logs, journals, goals and personal records. Its domain tools cover things such as habits, health and Korean study; an agent reads the relevant records and combines them for the task at hand. My macrostates describe conditions I'm working toward, so that history has a direction as well as a chronology.

The distinction between shared context and project work matters in daily use. A writing agent can recover an experience from my history, but the essay's drafts and illustration assets belong in its own folder. Decisions about the article stay with the article, where the next session can inspect them.

The interfaces I use around that context have different jobs:

Codex is my main work interface for research, coding, writing and study.
OpenCodex extends that setup to mobile and multiple AI accounts. Most of my work still happens locally in Codex.
Aside gives me an agent in the browser, with access to signed-in websites and browsing context.
Notion is a loose document workspace where I've tried writing processes and inspected intermediate material. Its freely arranged pages also inform what I'm looking for in an Idyllic interface.
Linear holds the work queues.

Job search and technical preparation

My background is in developer tools and programming languages. Before Idyllic, I worked in the Terra and CosmWasm ecosystem, building SDKs, simulators and CWScript. Agents help research roles and organize evidence of that work; I handle introductions and conversations.

The technical stage of interviewing makes me insecure because I haven't coded by hand in a while. Building through agents and demonstrating the same abilities unaided require different preparation. React fundamentals, TypeScript, debugging, data structures and algorithms are the areas I need to practice, alongside explaining the decisions in my previous projects.

Writing up those projects already gives me something useful for the search: public explanations I can point to. It also makes gaps in my own explanation apparent before I have to answer questions in an interview.

Micro-businesses and Busibody

A friend told me he'd reached roughly $1m in annual recurring revenue with a business that agents helped originate and run. That encouraged me to test a couple of stealth micro-businesses and learn more about the work around an offer.

My technical background doesn't answer the GTM questions automatically. Who needs this? What would make the offer understandable to them? Where would they encounter it? The experiments give those questions a specific offer to work on, with external examples and market observations to learn from.

Busibody provides shared infrastructure across the businesses. Each business has its own context, while common tools and records support the operating work. Needs that recur across the experiments inform what belongs in Busibody. I'm still directing the work; customer demand and revenue remain things to establish.

This is one connection to Idyllic: a business needs more than an agent capable of completing an isolated task. I have to supply the right context, inspect what happened and decide what follows. Building shared business tools gives me concrete coordination problems to work on.

Writing and public presence

My writing queue contains ideas from conversations, recorded walks and project work. For a selected piece, agents recover source material and help draft it. I choose the substance and review the result.

That process has produced published work. On September 11, I published Thinking in rhythms, How to write with AI and Mental models for working with AI on my site. The useful material comes from explaining an experience closely enough that someone else can understand the decision or mechanism involved.

Working with agents on those pieces also exposes review problems. A change to the argument can require new prose, while a request to fix one sentence should preserve the rest. When an agent loses that distinction, I have to reconstruct what we'd already agreed on. That is a specific requirement for Idyllic's treatment of context and review history.

My website, X and LinkedIn are the current publishing channels, with GitHub providing technical examples. YouTube and a collection of educational material for founders and agents remain possible extensions. The unresolved question is how useful public work leads to more conversations, collaborators and paid opportunities.

Languages and education

These interests have histories that predate the AI setup. Their current activity levels vary, and much of the structured study is still in the queue.

French

I first taught myself French, then spent a year at a French lycée in the literary Bac L track, studying philosophy and literature. I reached academic fluency and earned 88/100 on the DALF C2.

French became a language I read and thought in. I read deeply in philosophy and literature, and the French essay tradition influenced how I approach a question: identify the tension, then develop an argument that addresses it. That background is part of how I write now.

I'm still conversationally fluent, but rusty relative to my strongest period more than ten years ago. The aim is to recover ease, precision and range through subjects that hold my attention. Regular AI voice conversations about books and ideas are a proposed practice, not an established routine.

Korean

I've lived in Korea, and Korean is part of my family and community life. My grammar knowledge is ahead of my spontaneous speaking: understanding a construction doesn't mean I can retrieve it quickly in conversation.

A recent Korean church visit made the listening gap concrete. I enjoyed being there and meeting people, but couldn't follow the sermon. The language matters because it changes how directly I can participate in that community.

I have iTalki lesson material and vocabulary and grammar records in my Korean study tools. Those give a tutor specific material to work from. The practice I still need is listening and producing language in conversation, especially about everyday life and subjects I already discuss at length in English.

Mathematics and ML

I studied mathematics and computer science at Berkeley, including linear algebra, discrete mathematics and probability. My current study interests include proofs, probability and statistics, with information theory and cybernetics connected to my interest in systems.

Much of my technical experience comes from building languages, runtimes and developer tools. Building a language model from scratch is on my study list because it would connect the mathematics to the models underneath the tools I use. Working through implementations and derivations is the missing practice here.

My study and practice queue

These are candidates and materials to work from, rather than a claim that I'm taking all these courses:

Stanford CS336 and other LLMs-from-scratch material combine the theory and implementation work I'm looking for.
The "Modern Developer" series is also on my list. Coursera, MIT OpenCourseWare and Stanford are sources for structured study.
My DesignCode library holds references for React, interface design, animation and agent workflows. It already gives me examples to inspect when a project calls for an unfamiliar technique.
Coding quizzes and exercises could support interview preparation. For languages, I'm considering topic-based voice practice during walks or cardio.
Recurse Center, coding dojos and salons are possibilities for learning alongside other people.

For a course workspace, the unresolved interface question is concrete: after several days away, where do I see the last exercise, the part I struggled with and the next assignment? That belongs among Idyllic's design questions, even before I have a dependable study routine.

My body, my day and life with people

I've trained on and off for years and built a substantial strength base. Keeping that training present through changes in work and routine has been harder. My current goals are to get leaner while retaining muscle, improve bodyweight strength and build aerobic fitness.

A recent 35-minute elliptical session left me feeling grounded despite the resistance I felt beforehand. Walking and pickleball also give me movement I enjoy. Food at home and sleep support the same goals; activity and body records let me look back at what happened instead of relying entirely on how the week feels.

Being unemployed also means choosing more of the day's structure myself. Logs and check-ins help me see where the time went and return to priorities. I've also configured a weekly review.

I live in San Francisco with my wife. Shared meals, outings and household coordination belong in this arrangement too; Winha's Slack surface is one part of that shared work. Correspondence and practical documents need somewhere findable when a decision or follow-up comes up.

Church, sport and learning groups offer different kinds of contact with people. The Korean church visit connected language practice with community, and pickleball gave me both exercise and time with others. Those activities meet needs that an agent conversation doesn't cover.

What I'm building with Idyllic

The current Idyllic direction comes from managing this arrangement myself. Busibody has recurring business operations to support. Writing produces decisions about context and review. Study raises questions about returning to unfinished work. I still remember many of the connections and supply the next prompt.

I have a local workflow prototype that runs model steps and pauses for my feedback. It retains the inputs, outputs and review history for an execution. So far, I've tested that behavior on hypothetical material.

The design work has two main questions:

What structures can take over repeated coordination that currently depends on me prompting or relaying context?
What does the interface need to show so I can understand what happened, what it relied on and where my judgment is needed?

Agent journals are one requirement I'm considering. An agent helping with a business needs contact with the market; a tutor needs evidence of what I can do. Recording what it tried and learned would make that evidence available to the next agent, with less of the explanation passing through me.

I'm approaching the work from both directions. Participating in the activities supplies concrete problems, and examining their relationships helps me decide where shared infrastructure could reduce the work. I'm studying what becomes difficult as the arrangement grows more complex.

Eventually, I want other people to design their own setups and delegate more to an agent version of themselves. The questions I can work on now come from the system I'm using and the parts I still have to hold together.