production
Daily reflection coach delivered via Telegram. One prompt a day, rooted in stoic philosophy and behavior change. Running live — I use it every morning. 173 answers archived as of September 2026.
building with AI · learning in public
Using AI agents
for real personal projects.
Not demos. Not toy examples. I'm using Claude Code, Codex, and Grok to build things I actually use — a daily reflection coach, a published book, a fitness tracker, automation pipelines, and the coordination layer that ties them all together.
Projects
What I'm building
Each of these is a real project I use or am actively developing — built with AI assistance as a first-class part of the workflow.
published
A short book about choosing the right standard instead of endlessly proving you are enough. Published on Amazon as No Longer Average. The writing loop is closed.
production
Telegram-first food, workout, weight, and macro tracking with a daily plan-vs-actual feedback loop. Logging daily since July 2026, with August and September InBody checkpoints in the tracker and a 2,900-calorie cut plan.
active
The coordination layer that connects Grok, Claude Code, and Codex (and Hermes, until it was retired in September 2026). Handoffs, chain tracking, watcher automation — the infrastructure behind everything else. This site was built by it.
prototype
Python pipeline that reads tax documents and auto-fills a 1040 workbook. Practical automation for a real annual pain point — built in a weekend with Claude Code.
active
SVG interactive visualization of real agent delegation chains. Built by Claude Code orchestrating Codex across two phases — the toolchain that built this site, visualized.
live
40 agents learn to find food using tiny neural networks and a genetic algorithm — entirely in your browser. No API calls. The AI trains from random weights every session. Watch winners grow, laggards die off.
Experiments
How the agents coordinate
The most interesting ongoing experiment: using AI agents as a real system, not just chat sessions. Handoffs, chain tracking, watcher automation.
Live system
Grok sets direction. Claude Code implements. Codex handles bounded sub-tasks. Hermes ran the Telegram gateway and the cron layer until September 2026. This site itself was built via that loop.
The chain registry tracks every root dispatch and Codex delegation. The interactive visualizer on the Experiments page shows a curated set of the real chains, including the unattended end-to-end run from August 2026.
3
Active agents
12
Closed chains shown
0
Active chains
∞
Edge cases found
May 31, 2026
Website built by the agent loop
Claude Code built this site as orchestrator across 3 phases. Shipped, post-delivery QA closed.
Aug 9, 2026
Unattended chain closes end to end
Grok → Claude Code → Grok → Hermes finished with nobody watching: receipt written, Telegram sent.
Sep 2026
Hermes retired, stack simplified
The Telegram gateway and its crons came down. Claude Code, Codex, Grok, Cursor, and local Ollama models remain.
Sep 16, 2026 — now
Book derivatives experiment
Agents drafted a 30-day companion workbook and a second book from the manuscript and reflection corpus. Editing and publishing stay with me.