Copy My Claude Code Setup — The LearnAI Agent Stack From Scratch

This is the full reading path and setup checklist to replicate the LearnAI Claude Code environment on a new machine — from zero to a working agent harness with /goal, Codex review gate, oh-my-claudecode, and spec-driven workflows. Each layer builds on the last. Estimated time: 2-4 hours for a complete setup.

This hub collects the most important entries across this wiki into a single ordered path. Follow it top to bottom if you’re starting fresh, or jump to the layer that matches where you are.

The Stack at a Glance

Layer 0: Claude Code installed + API key
    ↓
Layer 1: CLAUDE.md — your global instructions (the most important file)
    ↓
Layer 2: oh-my-claudecode — orchestration plugin (/goal, /pause, agents)
    ↓
Layer 3: Skills — power tools for specific tasks
    ↓
Layer 4: Codex review gate — second-opinion AI that catches your mistakes
    ↓
Layer 5: Spec-driven workflow — spec.md → /goal → /pause → done
    ↓
Layer 6: Long-running sessions — orchestrator + headless pattern

Layer 0 — Foundation: Install & Mental Model

Time: 45 min

Start here if you’ve never used Claude Code or have only used it casually.

Install:

npm install -g @anthropic-ai/claude-code
claude

Layer 1 — CLAUDE.md: Your Global Instructions

Time: 30 min (reading + writing your own)

CLAUDE.md is the single most powerful file in your Claude Code setup. It defines how the AI behaves in every project — your defaults, your rules, your constraints. Everything else layers on top of it.

What belongs in CLAUDE.md:

• Your name and how you want to be addressed
• Global workflow rules (Codex review gate, git safety, test-first)
• Project-type-specific patterns (Next.js pitfalls, Docker notes, AWS SSH)
• Thinking-partner mode instructions

Example CLAUDE.md sections:

# User
Call me Q. Always.

# RULE #1 — Codex Review Gate
After completing ANY task, automatically send output to Codex for review.

# Git Safety
NEVER push to remote without asking first.

# How I Learn
Explain the "why" not just the "what". Draw ASCII diagrams for systems.

Read: SpecOps — Spec-Driven Development with AI Coding Agents — covers how to write specs that anchor your CLAUDE.md rules to specific tasks.

Layer 2 — oh-my-claudecode: The Orchestration Plugin

Time: 20 min

oh-my-claudecode adds the /goal, /pause, agent roles, and multi-AI orchestration. This is the scaffolding that turns Claude Code from a chat into an agent harness.

Install:

npm install -g oh-my-claudecode
omc setup

Layer 3 — Skills: Power Tools

Time: 30-60 min (install your top 10)

Skills extend Claude Code with domain-specific capabilities. The LearnAI stack uses:

Top skills to install immediately:

• oh-my-claudecode (Layer 2 — already done)
• pua (persistence/task discipline)
• prompt-master (prompt engineering assistant)
• grill-me (Socratic learning)
• review-wiki-skill (for wiki/writing workflows)

Layer 4 — Codex Review Gate: Your Second-Opinion AI

Time: 15 min (understand + wire into CLAUDE.md)

The Codex review gate is a mandatory second-opinion check after every significant output. Claude Code generates; Codex critiques. Together they catch mistakes neither catches alone.

Wire the review gate into CLAUDE.md:

# RULE #1 — Codex Review Gate
After completing ANY task, pipe output to Codex:
echo "CONTENT" | codex exec --model gpt-5.5 --ephemeral --color never -o /tmp/codex-out.md - >/dev/null 2>/dev/null && cat /tmp/codex-out.md

Layer 5 — Spec-Driven Workflow: Write spec.md First

Time: 10 min (install the CLI + run your first interview)

Spec-driven development means writing acceptance criteria before running /goal. The interview tool takes 5 minutes and produces a spec.md that anchors every turn of the agent session.

npm install -g @weihaoqu/project-spec-interviewer-skill
project-spec-interviewer-skill
# → spec.md written to current directory

/goal Complete all tasks defined in spec.md.
      Done when: [paste done condition]
      Scope: [files listed in spec.md]

Layer 6 — Session Management: Never Lose Context

Time: 10 min

Mac Migration: Clone This Setup to a New Machine

Pre-built export/bootstrap scripts handle the heavy lifting. Follow these steps in order — every command is copy-paste ready.

Step 1 — Export config from the old Mac

bash ~/.claude/claude-code-export.sh ~/Desktop/claude-export

Exports: CLAUDE.md, RTK.md, RESEARCH_WORKFLOW.md, settings.json, custom commands, agents, hooks, templates, memory index.

Step 1b — Export custom skills and memory content (gap in the export script)

Your custom skill .md files are NOT included in the export above. The same is true for the content of ~/.claude/memory/ — claude-code-export.sh copies the MEMORY.md index but not the *.md content files that index points at, so the new machine would have a memory index referencing files that don’t exist. Copy both separately:

mkdir -p ~/Desktop/claude-export/skills ~/Desktop/claude-export/memory

# Custom skill files you wrote (curriculum-designer, research-kb, etc.)
cp ~/.claude/skills/*.md ~/Desktop/claude-export/skills/ 2>/dev/null || echo "no .md skills"

# Global memory content files (the MEMORY.md index alone is not enough)
if [ -d ~/.claude/memory ]; then
  rsync -a ~/.claude/memory/ ~/Desktop/claude-export/memory/
else
  echo "no ~/.claude/memory directory — skipping"
fi

# If you want ALL skill packages too (~200 dirs, potentially large):
# rsync -av --exclude='*.pyc' ~/.claude/skills/ ~/Desktop/claude-export/skills/

On the new Mac, after Step 8 restores skills, also restore the memory content. Use whichever transfer path matches Step 2:

mkdir -p ~/.claude/memory

# If transferred via Dropbox:
rsync -a ~/Dropbox/claude-export/memory/ ~/.claude/memory/

# If transferred via rsync/AirDrop:
rsync -a ~/claude-export/memory/ ~/.claude/memory/

Re-running the export? claude-code-export.sh (Step 1) removes and recreates the export directory when it detects a prior export. That wipes anything Step 1b put there too. If you re-run Step 1, re-run Step 1b right after — don’t assume the prior memory/skill copy is still in the bundle.

Step 2 — Transfer to the new Mac

# Option A: Via Dropbox (recommended — syncs automatically)
cp -r ~/Desktop/claude-export ~/Dropbox/claude-export
# Wait for Dropbox to sync, then continue on the new Mac
# ⚠️ Dropbox Smart Sync warning: files may show in Finder but not be downloaded yet.
# Force download on the new Mac before running the bootstrap:
#   brctl download ~/Dropbox/claude-export  (if using iCloud-style Dropbox)
#   or: open ~/Dropbox/claude-export in Finder → select all → right-click → "Make Available Offline"

# Option B: Via rsync over SSH (both Macs on same network)
rsync -av ~/Desktop/claude-export/ USERNAME@NEW-MAC-IP:~/claude-export/

# Option C: AirDrop — Finder → right-click ~/Desktop/claude-export → Share → AirDrop

Step 3 — Install prerequisites on the new Mac

# Install Homebrew (macOS package manager — skip if already installed)
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

# Install core tools
brew install node git jq gh ffmpeg

Step 4 — Install Claude Code

npm install -g @anthropic-ai/claude-code

Step 5 — Run the bootstrap script

# If transferred via Dropbox:
bash ~/Dropbox/claude-export/claude-code-bootstrap.sh ~/Dropbox/claude-export

# If transferred via rsync/AirDrop:
bash ~/claude-export/claude-code-bootstrap.sh ~/claude-export

The bootstrap script:

• Copies CLAUDE.md, settings.json, hooks, commands, agents into ~/.claude/
• Rewrites absolute paths from /Users/oreo to your new $HOME automatically
• Prints all /plugin install commands at the end — copy and run those next

Step 6 — Install plugins (from bootstrap output)

Open Claude Code and run the plugin commands the bootstrap script printed. They look like:

# In Claude Code:
/plugin marketplace add openai/codex-plugin-cc
/plugin marketplace add jimliu/baoyu-skills
/plugin marketplace add Lum1104/Understand-Anything
/plugin marketplace add kepano/obsidian-skills
/plugin marketplace add lingfengQAQ/webnovel-writer

/plugin install ai-generation-skills@baoyu-skills
/plugin install utility-skills@baoyu-skills
/plugin install obsidian@obsidian-skills
/plugin install codex@openai-codex
/plugin install understand-anything@understand-anything
/plugin install webnovel-writer@webnovel-writer-marketplace

/reload-plugins

Step 7 — Install npm tools + RTK

# oh-my-claudecode (orchestration layer)
npm install -g oh-my-claudecode
omc setup

# Spec interview tool
npm install -g @weihaoqu/project-spec-interviewer-skill

# Codex CLI
npm install -g @openai/codex

# RTK (token-saving proxy)
brew install rust && cargo install rtk
# or: download binary from https://github.com/rust-lang/rtk/releases

Step 8 — Restore custom skills

# Copy the .md skill files exported in Step 1b
cp ~/Dropbox/claude-export/skills/*.md ~/.claude/skills/ 2>/dev/null || true

# If you also exported full skill packages, restore them:
# rsync -av ~/Dropbox/claude-export/skills/ ~/.claude/skills/

Step 8b — Migrate the progress.md handoff skills

If you use the progress.md Handoff workflow, there are two halves that need to travel:

Half 1 — User-level skill directories that back the handoff workflow (one-time per machine). These are SKILL.md files inside named subdirectories (~/.codex/skills/checkpoint/SKILL.md, etc.), NOT top-level .md files — Step 1b’s *.md glob misses them. Copy the whole directories explicitly:

# On the old Mac — add to your export
mkdir -p ~/Desktop/claude-export/skills-tree/{codex,claude}
cp -R ~/.codex/skills/checkpoint  ~/Desktop/claude-export/skills-tree/codex/
cp -R ~/.codex/skills/resume      ~/Desktop/claude-export/skills-tree/codex/
cp -R ~/.claude/skills/checkpoint ~/Desktop/claude-export/skills-tree/claude/
cp -R ~/.claude/skills/resume     ~/Desktop/claude-export/skills-tree/claude/

# On the new Mac — restore (after Dropbox/rsync transfer)
mkdir -p ~/.codex/skills ~/.claude/skills
cp -R ~/Dropbox/claude-export/skills-tree/codex/*  ~/.codex/skills/
cp -R ~/Dropbox/claude-export/skills-tree/claude/* ~/.claude/skills/

After this, in Claude Code the /checkpoint and /resume slash commands resolve in any project on the new Mac that also has the per-project handoff files from Half 2 — both halves are required. In Codex the same workflow is reachable by typing the bare trigger words checkpoint / resume (no slash; Codex doesn’t honor user-defined slash commands — the skill activates on those trigger phrases from its description and AGENTS.md).

Half 2 — Per-project handoff files (already in git for projects that have it). The script scripts/agent-handoff.sh, the project-level Claude Code slash command files .claude/commands/checkpoint.md and .claude/commands/resume.md, AGENTS.md, and the live progress.md log all live inside each repo and are tracked. For projects that already have the workflow, they come with git clone — nothing to install. For a NEW project (or one you didn’t git clone), use the one-command installer:

cd /path/to/new/project
bash ~/.claude/skills/checkpoint/install.sh

That copies the 4 files (script + 2 slash commands + AGENTS.md if absent) and adds two .gitignore entries. Idempotent and safe to re-run. Existing AGENTS.md is never overwritten — if you have one, the installer tells you what to append manually. ~/.codex/skills/checkpoint/install.sh is the mirror for running from Codex. Source path defaults to ~/Desktop/learnai-course (the canonical repo); override with PROGRESS_HANDOFF_SRC=/path/to/source if you keep it elsewhere.

progress.md is NOT copied by the installer — it’s created on your first /checkpoint in that project.

(Note on terminology: .claude/commands/checkpoint.md is a Claude Code project-level slash command (gives you /checkpoint inside that project). ~/.claude/skills/checkpoint/SKILL.md is a Claude Code user-level skill (also reachable as /checkpoint from any project — the cross-project fallback). ~/.codex/skills/checkpoint/SKILL.md is a Codex skill — but Codex doesn’t expose user skills as slash commands, it activates them on trigger phrases like checkpoint, save context, wrap up listed in the skill’s description and in AGENTS.md. So the Claude-side install gives you a slash command; the Codex-side install gives you a triggered skill. Both invoke the same scripts/agent-handoff.sh.)

The wiki entry has a Quick Reference section if you forget the daily-use commands: https://weihaoqu.github.io/learnAIDoc/wiki/progress-md-codex-handoff/

Step 9 — Authenticate Codex

codex login

Step 10 — Let Dropbox sync your projects

Install Dropbox on the new Mac, sign in, and wait for ~/Dropbox/ to sync. Your wiki (learnAIDoc/), research KB, and all project files appear automatically — no manual copying needed.

Post-Bootstrap Checklist

# Verify RTK works
rtk gain

# Check oh-my-claudecode health
omc doctor

# Verify hooks fire (run any git command in Claude Code)
git status

# Update API keys file with new machine's keys
open ~/.claude/api_keys.md

Then open Claude Code and:

• Run /goal with a small test task (confirms /goal + hooks work)
• Run project-spec-interviewer-skill (confirms npm tool works)
• Test one Codex command: echo "hello" | codex exec --model gpt-5.5 --ephemeral -
• Set up MCP servers (Google Calendar, Gmail, Drive) if needed

What is NOT migrated by the export script

How LearnAI Team Could Use This

• Onboarding new collaborators — Send this page instead of a 2-hour setup call. New team members follow Layer 0 → 6 in order and arrive with the same environment.
• CS-310 / CS-336 student setup — Assign Layer 0-2 as Week 1 homework (Claude Code + oh-my-claudecode + first /goal session). Grade the spec.md they produce, not the code.
• Lab computer setup — When setting up a shared lab machine, use the bootstrap script and this reading path. The CLAUDE.md can be customized for research context (security analysis, type systems, etc.).
• Teaching harness engineering — Use the Layer 4 (Codex gate) + Layer 5 (spec-driven) combination as a “responsible AI workflow” module. Show students how the harness catches mistakes the model itself misses.

Real-World Use Cases

Important Things to Know

• CLAUDE.md is the highest-leverage file. The model, the skills, the plugins — these matter. But CLAUDE.md is where you define the rules that make all of them work together. Invest 30 minutes writing it carefully; it compounds across every session.
• Layer order matters. Don’t install 20 skills before you understand /goal. Don’t run /goal before you have a spec.md. The layers are ordered by dependency — each one makes the next more valuable.
• The export script captures config, not state. Skills, plugin caches, and session history are not exported. The new machine will feel slightly “fresh” — that’s intentional.
• Dropbox handles the content. If your projects, wiki, and research KB live in Dropbox, migration is just the ~/.claude/ config transfer. The content is already synced.
• Start with one /goal session. After Layer 2, don’t wait until you’ve read all layers. Run a real task through /goal with a spec.md. The experience teaches more than reading.