OPERATING
MODE.

This is May 2026. I live in a terminal window all day every day. 70 to 80 percent of what I do on any of my machines lives in there. Claude Code is the common denominator across everything I build, and the operating mode below is the substrate I built around it once context survival became the actual bottleneck.

Cursor lost me in June 2025. I unsubscribed, churned out, and stuck with Claude Code. I had been a vim, zsh, bash, jq user for years, and I had built developer platforms in product roles where the IDE was the constant across a long career. Claude Code outran every IDE I had been using, and the unsubscribe was a natural outcome.

July 2025 is when I went ground-up on Paxora. The security model, the backend, the API surface, and the agent pipeline that turns uploaded documents into court-ready California family-law forms all came up out of Claude Code from the first commit. I bootstrapped the company alone.

[CONTEXT BECAME THE BOTTLENECK]

The early months were vibe coding the hell out of everything. Compaction would fire mid-thought. I would explain something, watch the model forget it two prompts later, and start over. Oh, I just explained this to you. Why do I have to tell you again? CLAUDE.md was already in the repo and it was not pulling enough weight on its own. Context was the actual bottleneck, and I had been treating it like an afterthought.

So I started writing things down. Naming conventions formed organically across research, spec, plan, audit, analysis, guide, state. Each journal time-stamped, each one pinned to a specific decision so the next session had something to attach to after compaction fired. Nobody taught me this. The habit formed because the alternative was re-explaining the same context every morning.

[58 AGENTS]

That fall, I had 58 agents running at one point, one per California county, codifying jurisdiction rules for family law. 58 context windows that needed to write into independent journals so the orchestrator did not choke. The orchestrator only saw synthesis. Each county did discovery in its own window. The whole pattern crystallized there.

Between July and December 2025, the platform evolved alongside the practice. The model lineup expanded across multiple Sonnet and Opus versions; skills, agents, and hooks landed inside Claude Code; plugins arrived; and several open-source skills came to life. My own habits stratified into a workflow alongside that arc.

February 2026 was the moment the personal side reached the professional side. I bought a Mac Mini, set it up as Hermes, and ran OpenClaw on it for local inference. Bodhi, my personal agent persona, runs on top of Hermes through SOUL.md with the same context discipline I use for work. The morning brief reading my health data lands on Telegram before I open anything else. The same operating mode that built Paxora and the fractional engagement is the substrate I am extending into my own life.

[A DAY IN THE STACK]

Here is what a day inside that operating mode looks like in May 2026.

A typical Tuesday on the personal laptop runs four sessions. The first tab carries Paxora infrastructure with the Docker daemon, the front end, the backend, and debug logs streaming. No Claude session lives there. The whole tab is real estate for watching the system, restarting it, hot reloading, and spotting errors. Claude mounts on the Paxora repo in the second, handling product engineering for whichever shape the day demands: a bug fix, a feature enhancement, a design-partner request. Outreach, CRM, and analytics run in the third tab, probing how the email cohorts perform and what to move next. The fourth tab floats. Some afternoons it carries ad-hoc scripting against a new open-source tool I want to understand. Lately it tends to be SSH into the Mac Mini to tune Bodhi's instructions while my morning brief composes.

Switch to the fractional laptop and the shape changes but the discipline does not. Two or three sessions running product engineering, eval work, and interface design for agentic systems. The third session, when it is open, holds an investor deck or a conference demo or a piece of marketing collateral in progress.

Underneath all of this sits a cross-project skill layer cloned on every machine and symlinked as global Claude context. Project-specific .claude folders live local per repo, with dozens of skills across both engagements handling specialized work and standing devil's-advocate sub-agents pushing back so I do not drift. Hooks gate new file creation and review passes at the repo level. They handle the line-level review I would otherwise be doing myself. I rarely read the code line by line. That altitude comes from upstream work: deliberate discovery, system design, and the spec the model executes against, all in place before any code is written. I track file structure, naming conventions, router patterns, the authentication flow, and consistency across modules. The agents and hooks watch the line work; I work at that altitude because that is where the leverage sits. The premise underneath all of it is that I am alone, and I needed an endless set of experts watching the side and the back.

[CROSS-POLLINATION · BOUNDARIES]

Cross-pollination stays at the tooling layer. Auth0 is the example I name out loud, the same integration pattern reused across both sides through a journal tagged across folders. Domain context never crosses engagements. Each side keeps its own analysis lane. The compounding within domain is where the real benefit lives, and the persistent context I carry across machines and sessions gives me the confidence to bootstrap anything inside either domain quickly.

The substrate is a multi-machine layer engineered for context survival through symlinks, Git, and a shared drive. Concurrent operation across Paxora, the fractional engagement, and personal automation rides on top of it, with journals as the deliberate discipline that makes the whole thing survivable.

One standing instruction sits underneath every Claude Code instance and every sub-agent I dispatch. It lives in CLAUDE.md and SOUL.md, propagates through the skills layer, and runs across every agent in the fleet. Inherited patterns are treated as a hypothesis until they survive a teardown.

The conversation that produced this essay lives in time-stamped journals that survive compaction.

[FORWARD]

Over the next twelve months I expect personal AI assistance to commoditize across watches, wearables, and every richer signal one level closer to brain, nervous system, and heartbeat. Attention will get scarcer. The work between now and then is to keep this operating mode running on its own and hand more of the line work to the agents each cycle.