A working cookbook for running Claude Code as a system, not a chat box.
The model is the commodity. The harness is the moat.
This repo is the wiring diagram behind one operator's full Claude Code setup: the rules, skills, hooks, memory, learning loop, session search, second brain, multi-provider routing, scheduled jobs, and always-on agents, plus the reasoning for why each piece exists and how they connect.
It is not a catalog. The community already maintains one of those at awesome-claude-code, and you should use it: that is where you go to find a skill, a hook, a slash command, an MCP server. Claudelicious is the other half. It shows how those pieces fit together into a harness that remembers across sessions, improves itself when you correct it, and runs while you sleep, all wired around a homelab.
Most "here is my setup" repos are a flat list. This one leads with the systems underneath the skills, because those are the part you cannot reverse-engineer from a screenshot.
If a human pointed their Claude Code at this repo to help build or understand a harness: start at AGENTS.md. It tells you how the repo is laid out, the conventions every doc follows, and how to adapt a pattern without copying anyone's private paths. Then read docs/00-the-map.md.
Every doc is written to be applied, not just read: principle first, then a worked example, then explicit ship-or-scrub notes.
| You want to... | Go to |
|---|---|
| Get the spine in place in your first week | QUICKSTART.md |
| Read the whole thing as one long story | STORY.md |
| Understand the whole thing in 10 minutes | 00 The map + Philosophy |
| Stop repeating instructions every session | 01 Rules and context |
| Build a skill library that stays small and sharp | 02 Skills |
| Make the agent stop doing dangerous things | 03 Hooks |
| Fix that Claude does not know the date | 03 Hooks |
| Make Claude remember things across sessions | 04 Memory |
| Make a correction stick instead of recurring | 05 The learning loop |
| Resume work cleanly after a compaction | 06 Continuity |
| Search your own past sessions | 07 Session search (mneme) |
| Let Claude semantically search your notes | 08 The second brain |
| Run Claude Code on Kimi / MiniMax / GLM / Qwen | 09 Multi-provider |
| Schedule jobs inside Claude Code and on the machine | 10 Crons and scheduling |
| Build an agent that runs on a heartbeat | 11 Always-on agents |
| Make the harness write in your voice | 12 Voice and anti-slop |
| Generate logos, infographics, diagrams | 13 Visual generation |
| Fork work to subagents, or let the harness orchestrate itself | 14 Subagents and workflows |
| Run a loop until the work is done, not just on a timer | 15 Loops and autonomy |
| Spread the harness across more than one machine | 16 The fleet |
Tune settings.json so a session behaves the way you want |
17 Settings |
| Connect the harness to mail, calendar, search, a browser | 18 MCP |
| Understand running with broad machine access, and the guards | 19 Running wide open |
Templates and copy-able files live in templates/, hooks/, skills/, and settings/.
Everything here is built from five primitives. Lock these first.
| Word | What it is |
|---|---|
| Model | The LLM. The smallest decision you make. |
| Harness | Everything around the model. The part that compounds. |
| Agent | A model in a harness, given a goal and the room to pursue it. |
| Skill | A repeatable procedure the harness invokes by name. A business process, encoded. |
| MCP | Model Context Protocol. How the harness reaches external systems. |
A great model with no harness is a demo. A harness with a great model is a second nervous system.
To make the scale concrete: a skill library deliberately pruned from a peak of over 130 down to a curated few dozen, because the cut is the craft and not a footnote to it. Around 15 MCP servers, each a doorway into a real system. A four-tier memory taxonomy over roughly seventy thousand plain-markdown vault documents. A mesh of five machines from a laptop to a GPU box. Eight named agents, each with its own role and memory, for less than a couple of lunches a month.
None of those numbers is the point. That they are governed is the point. A system this size that nobody pruned would collapse under its own context.
Not a clone-and-run starter kit. It is docs-first, with templates and worked examples so you adapt the pieces to your own paths, hosts, and vault. The reasoning is the product. A scaffold you paste in without understanding becomes the next thing you cannot maintain.
It is also scrubbed. No real hostnames, IP addresses, credentials, or private paths. Where the original used a specific path or host, the docs use a placeholder and tell you what to substitute.
This cookbook is the implementation. The reasoning for why you would build a harness at all, what it does to your working life, and how to lead a team across the same gap, lives in the book: Builder-Leader: The AI Exoskeleton That Crosses the Gap (builder-leader.com).
The book is the why. This is the how.
The ideas keep developing in public on two blogs: AIXplore, the technical, practitioner-facing one closest to this repo, and Run Data Run, the leadership and strategy angle. A lot of these patterns were first worked out in the open there.
Dual-licensed by content type. The code you would copy into your own setup —
everything in templates/, hooks/, and skills/ — is MIT (LICENSE),
so use it however you like. The prose — docs/, this README, and PHILOSOPHY.md —
is CC BY-NC 4.0 (LICENSE-docs): share and adapt it with
attribution, but not for commercial repackaging.
Start small. Be a builder.