A memory system, two macOS apps, a remastered EP, and a 5-chapter writing series—none of it planned.
In the last five months, I shipped more than in any year of my career. Not because I was grinding. Because nobody was telling me what to work on.
I was laid off from Lyft in October 2025. I'd been there as a Learning Architect—building programs, automation, and AI enablement for 4,000+ employees. When the structure disappeared, the same instincts that built learning systems started building other things.
The First Thing: A Memory
In early February 2026, I started a project I didn't know was a project.
I was frustrated with Claude. Not the model—the amnesia. Every session started from zero. I'd explain my background, my preferences, my working style, my projects—and the next day, none of it existed. It was like having a brilliant colleague who got a complete memory wipe every night.
So I built a workaround. A folder of markdown files that Claude reads at the start of every session: who I am, what I'm working on, what I've already decided, what I've corrected. Plain text. Git-tracked. No database. No app.
I called it stonerOS.
Within two weeks, it had a three-layer memory architecture and a corrections system. Within a month, it had grown into something I hadn't planned—dozens of specialized agents, automation hooks, scheduled tasks running while I slept.
I didn't design this upfront. The corrections file exists because the AI confidently told a recruiter I'd built for 5,000 employees when the real number was 4,000. The agent system exists because I got tired of loading the same context for different tasks. The safety layer exists because I accidentally let an autonomous agent write to a protected file at 3 AM.
Each piece emerged because the previous piece needed it.
The Second Thing: Music
In December 2025, two months after the layoff, I made an EP.
AI tools handled the production—instruments, vocals, engineering. I wrote every word, made every structural choice, and picked which of 20+ generations to keep for each track. The same ear I use for design work: knowing when something lands and when it doesn't.
The result—THE EXIT / Parallels—is a 7-track duet about two perspectives on the same relationship. I later remastered it with a custom Python pipeline: EQ, compression, loudness normalization, multi-format export.
I published it on SoundCloud and told no one for three months.
The Third Thing: Apps
In March, stonerOS was mature enough that I started noticing other problems worth solving.
ClaudeUsage started as a question: how much is this AI subscription actually worth? I built a native macOS menu bar app that tracks Claude's API-equivalent token cost in real time. When I checked the first month's data, it showed over $3,000 in compute value against a $20 subscription—numbers the app calculates by mapping local conversation files to current API pricing. OAuth from Keychain. No backend.
Drift started as a different question: why do I keep falling into the same work patterns? I built a macOS app that monitors my work sessions for behavioral signals—scope creep, infrastructure-first rabbit holes, boredom-driven pivots—and surfaces gentle nudges before I'm too deep. Built it in a single day. CI/CD pipeline, notarized, GitHub release.
Neither app was on a roadmap. They came from paying attention to my own friction.
The Fourth Thing: Writing
In mid-March, I started a 5-chapter series called I Gave Claude a Brain. Not a tutorial—a practitioner's narrative about what happens when you give AI actual memory and structure.
Chapter 0: what I built and what surprised me. Chapter 1: why memory changes the AI relationship. Chapter 2: the three-layer architecture and why plain files beat databases. Chapter 3: the 21-agent system and how I allocate work across model tiers. Chapter 4: what it means—the gap between specifying what you want and recognizing quality when you see it.
Then standalone pieces: What My AI Taught Me About Myself (pointing stonerOS at my own iCloud data and seeing behavioral patterns I hadn't named). The Learning Reset (why corporate AI enablement is stuck at the cheatsheet stage). The EP write-up. A piece about thread-pulling—how a Monday afternoon of discount-hunting turned into a portfolio overhaul, a remastered EP, and a content calendar.
None of it planned. All of it connected.
The Pattern
Here's what I notice when I look at the last five months:
Every significant thing I built started as a specific frustration, not a goal. stonerOS started because Claude kept forgetting me. ClaudeUsage started because I wanted to know if the subscription was worth it. Drift started because I kept scope-creeping. The EP started because I had words and no way to make them audible. The writing started because I'd built things worth explaining.
The thread-pulling isn't distraction. It's architecture. Each output creates a question that leads to the next build. The sequence is nonlinear, but the connections are real.
What I Notice
I followed my curiosity through a stretch where nobody was telling me what to work on.
At Lyft, I was a learning architect—which meant I built programs, automation, and enablement within that scope. On my own, the same instincts produced a memory system, two native apps, a music project, and a writing practice. The skillset didn't change. The constraints did.
Josh Stoner is a Learning Architect and systems builder based in Brooklyn. He spent 10+ years building learning programs at scale—most recently at Lyft—and now builds AI memory systems, native macOS apps, and writes about what happens when you give AI actual context. He publishes at josh-stoner.github.io.