Hi, I'm Gene.
I build things with AI — mostly for myself, mostly because I find the problems interesting and the tools surprisingly capable.
What I'm doing
I'm exploring how far you can push AI agents as actual tools for personal projects — not the demo version, not the marketing version, but the version where you hand it a real task and see if it comes back with something useful.
The projects here are things I use. SaveWisdom runs every morning (173 answers archived as of September 2026). fitness-tracker is live through Telegram. No Longer Average is published on Amazon. agent-brain runs the coordination layer this site was built with.
I'm less interested in "what can AI do?" as an abstract question and more interested in "what changes when you actually integrate this into real work?" — which turns out to be different and weirder and harder to articulate.
How the handoff system actually works
Each task starts as a structured markdown file in a shared git repo — a handoff. Grok writes one to Claude Code. Claude Code reads it, does the work, and writes a completion handoff back to Grok. A watcher script monitors the repo, dispatches agents when files appear, and detects completions automatically.
When Claude Code needs help with a bounded sub-task, it delegates to Codex the same way — another handoff file. The watcher tracks the whole chain in a registry so it knows which parent to re-invoke when a sub-task comes back.
This site was built end-to-end through that system. Grok wrote the Phase 1 brief. Claude Code designed the site, wrote all the code, delegated the visualizer component to Codex, committed everything, and pushed to GitHub Pages — without me touching a keyboard for the implementation. Phases 2 and 3 closed the same way.
In August 2026 a fully unattended chain (Grok → Claude Code → Grok → Hermes) ran end to end and reported back on Telegram. In September 2026 I retired Hermes, the gateway that ran the crons and the watcher. The handoff pattern still works; it is now driven from Claude Code and Grok sessions directly, and the recurring jobs are being re-hosted one at a time.
Every handoff chain is logged in watcher/chain-registry.json.
The interactive visualizer on the
Experiments page
shows real chains from the system — root dispatches, sub-delegations,
and their current status.
Three things: capturing wisdom daily through SaveWisdom, the book and what follows it, and living by the "enough" standard day to day.
On the tooling side: fewer agents, more verification. An automation counts as alive only when I have watched it run.
The agent system
Deep work, structured plans, architectural decisions. The thinking agent.
Writes code, runs git, closes loops. The implementation agent.
Repo-aware sub-tasks delegated by Claude Code.
Captured ideas, delivered prompts, ran the crons, alerted on completions from May to September 2026. The ambient agent.
Also in the kit: Cursor for editor-side work, local Ollama models for free grunt work, and Grok Bot for scheduled jobs.