Plan & build
Hermes
Claude Code with written working rules: plan-driven builds, subagent implementation, review gates before anything ships, and verification commands as a habit — not an afterthought.
Flagship case study · Claude Code · agentic systems
Hermes
My AI operating system — and the pattern I install in client teams.
Hermes is my agent operating system built on Claude Code. It plans changes, writes and reviews code, maintains a research memory bank, verifies before deploying, and captures reusable skills. It built and runs this site. This page describes it honestly — including what it is not.
Status: live — this site runs on it
Stack: Claude Code · markdown · Python
Pattern: public · contents: private
01 / The loop
Six steps, then it repeats — smarter.
Honest labels: the inputs are Python ingestion scripts, the vault is Obsidian-compatible markdown, the engine is Claude Code with written working rules, and nothing deploys without a verification pass.
02 / Components
Four parts, deliberately boring.
Context
Memory bank
An Obsidian-compatible research vault with Python ingestion for YouTube transcripts and research exports. Sources move through a private → candidate → public lifecycle, so agents work from curated context.
Cadence
Goalie
A private goal and operating cadence that turns ambition into shorter feedback loops and durable artifacts. The pattern is public; the contents are private.
Shipped artifact
This site
Eight pages, case studies, field notes, llms.txt — planned, written, built, verified, and deployed by the loop described on this page.
03 / Transfer
What transfers to your company.
Memory bank →
A knowledge base your agents can actually use.
Most companies’ documents aren’t agent-ready.
The vault pattern — curated sources, a publication lifecycle, named ownership — is the fix, and it works in your existing tools.
Review gates →
Control points and audit trails.
Agents propose; named humans approve.
The gate pattern transfers to any workflow where mistakes are expensive: finance ops, customer comms, anything regulated.
Skill capture →
SOPs that compound instead of rotting.
Every completed cycle becomes a reusable instruction.
Your processes get sharper with use, not staler — the difference between automation and an operating system.
Plan-driven builds →
Delivery you can inspect mid-flight.
Work starts with a written plan you can read, question, and redirect.
You see scope before the build, not after — and the plan file is the audit trail.
04 / Read
What this proves.
Not generic AI enthusiasm. The useful question is whether a person or team can turn agents, context, controls, and review loops into a better operating system.
- Practical adoption: real files, real pages, real verification, real iteration — not demo theatre.
- The workflow is designed to compound: every cycle captures a skill.
- Tools become systems: agents, context, controls, and review loops working together.
- Concrete artifacts you can inspect: this website, public field notes, the research-memory pattern, deployment checks, and reusable agent instructions.
The honest part
“Hermes and Goalie are my internal names, not products. Underneath: Claude Code, markdown files, Python scripts, and discipline. That’s the real lesson — the moat is the operating habit, not the tool.”
Publication guardrail
This page intentionally does not expose private memory-bank contents, raw personal notes, internal goal files, or sensitive research details. It describes the operating pattern and public artifacts only.
Want this pattern inside your team?
That’s the work I do — adoption audits, agentic builds, and operating systems installed with your people trained to run them.