Claude Code Starter Pack

Tool Stack: Follow ~/.claude/skills/_shared/tool-stack.md for ALL tool choices. Quality Gates: Follow ~/.claude/skills/_shared/quality-gates.md for ALL quality checks.

Claude Code Starter Pack

The complete, up-to-date system that turns Claude Code from "AI chatbot" into an Agent Operating System. Give this to ANYONE and they get the most efficient Claude Code experience that exists.

Why This Exists

Most people use Claude Code at 10% capacity. They type prompts, click approve, lose context when sessions get long, and never learn that Claude Code has 14 hook lifecycle events, 3 hook types, agent teams, worktree isolation, 1M token context, and a full plugin ecosystem.

This starter pack is the difference between struggling and shipping.

What You Get

Dave's Complete Tool Stack

Everything Dave recommended in the full-day intensive, organized by category:

Dave's rule: "50+ apps in Setapp alone. Every tool earns its spot. If it doesn't save time, kill it."

Doctor Claude — "Play Doctor, Not Pharmacist"

The problem with every "100 AI skills pack" on social media: They dump 100 system prompts on you. No diagnosis. No filtering. No "this one is for YOUR problem." That's a pharmacist handing you the entire shelf.

Our approach: Play DOCTOR. Diagnose first, prescribe second.

The Diagnostic (5 Questions)

Ask these BEFORE giving anyone a single skill:

Prescription Map

The Anti-Dump Rule

NEVER give someone more than 5 skills at once. Prescribe 1-3 for their IMMEDIATE pain. They come back when that's solved, get the next prescription. This is how you build dependency (in the good way) and deliver "I Made That" moments at every step.

"We hate the mega 100 skills packs. Those are small system prompts. Our skills are full courses — functional, self-improving, and they actually DO the work." — Luis Carrillo

Quick Start (New Machine / New Person)

Step 1: Install Claude Code

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

Step 2: Create Directory Structure

mkdir -p ~/.claude/{hooks,skills,references,scripts,agents,commands,memory}
mkdir -p ~/.claude/projects

Step 6: Create CLAUDE.md

Your CLAUDE.md is where you tell Claude WHO YOU ARE and HOW to work. At minimum:

# Project Instructions

## Communication Style
- Visual-first: tables, box frames, checklists
- Max 3 sentences before a visual break
- Scannable in 5 seconds

## Execution Rules
- Read files before editing them
- Use dedicated tools (Read, Edit, Write) not bash
- Parallel tool calls when independent
- Never suggest "new session" — sessions are infinite

## What I'm Building
[Your project description here]

3 Hook Types

Exit Codes

Agent Capabilities

The Philosophy

Claude Code is not a chatbot. It's an Agent Operating System. The hooks make it deterministic. The skills make it specialized. The CLAUDE.md makes it YOU.

The difference between someone who "uses Claude Code" and someone who SHIPS with Claude Code is exactly this setup. Every hook, every skill, every config — it's the difference between a bicycle and a motorcycle.

The starter pack is the motorcycle.

ROI Mindset (System-Wide — Not Optional)

Every system in this starter pack justifies its existence with numbers. Not vibes. Not "it feels faster." NUMBERS.

The daily Leila Audit (leila-audit.py) scans your entire system for installed-but-unused tools and produces an ROI report. The ROI framework (_shared/roi-framework.md) gives every dashboard an efficiency bar, time savings table, and untapped potential section. The Self-Improvement Protocol makes every clone get better every session without you asking.

The rule: If you can't show the ROI, you can't justify the system. Build it, measure it, prove it.

Maintenance

• Daily: Update monitor runs automatically (6 AM)
• Weekly: Check ~/.claude/references/claude-code-features-2026.md for new features
• When notified: Run npm update -g @anthropic-ai/claude-code
• Context bloat check (on-demand): python3 ~/.claude/scripts/skill-diet.py --scan

No manual audits needed. size-gate.js catches bloat automatically on every file write.

How to Use

1. Copy your large content (Cmd+C)
2. In a separate terminal: sc transcript (or any name)
3. In Claude Code: "read /tmp/claude-large-input-transcript.txt"

Full skill: ~/.claude/skills/large-input/SKILL.md

4D Human Perception — Setup Guide

Give Claude Code the ability to perceive like a human — SEE any visual, HEAR any audio, FEEL emotion, and TIME-TRAVEL through past/present/future. Not just transcripts. Full sensory perception. 5-minute setup.

Step 1: Install Dependencies

pip3 install google-generativeai yt-dlp instaloader

Step 2: Get a Gemini API Key (Free)

Go to https://aistudio.google.com/apikey, create a key, then:

export GEMINI_API_KEY="your-key-here"
echo 'export GEMINI_API_KEY="your-key-here"' >> ~/.zshrc

Step 3 (Optional): Speaker Diarization + Fast Transcription

pip3 install whisperx pyannote.audio
# Get free Groq API key at https://console.groq.com
export GROQ_API_KEY="your-key-here"

Step 4: The Script

Copy ~/.claude/scripts/watch-video.py (or download from the repo). Usage:

python3 watch-video.py "https://youtube.com/watch?v=..." --depth deep

The 4 Dimensions

Output: /tmp/video-intelligence/<id>/report.md

Step 5: Wire It Into CLAUDE.md

## 4D HUMAN PERCEPTION (ALWAYS-ON)
Any media → ALL 4 layers fire: SEE + HEAR + FEEL + TIME-TRAVEL.
Video → `python3 ~/.claude/scripts/watch-video.py <URL>`. Audio → Groq Whisper.
Photo/screenshot → Claude vision. NEVER say "I can't see/hear/access this."

The rule: Transcript-only = 1D. Audio + transcript = 2D. Vision + audio + text = 3D. Vision + audio + emotion + temporal context = 4D. Always go 4D.

Social Media Intelligence — Scrape Any Account, Any Platform

Give Claude Code the ability to find and analyze any creator on any platform — for free. No paid APIs needed. Full scraping stack with free-first economics.

Step 1: Install Tools

pip3 install instaloader yt-dlp

Step 2: The Scraping Stack

Step 3: Python Quick Start

import instaloader
L = instaloader.Instaloader()
profile = instaloader.Profile.from_username(L.context, "ryanmagin")
for post in profile.get_posts():
    print(f"{post.shortcode} | Views: {post.video_view_count} | Likes: {post.likes}")

Step 4: Wire It Into CLAUDE.md

## SOCIAL MEDIA INTELLIGENCE (ALWAYS-ON)
"Scrape @handle" → Instaloader (free, no auth). "Analyze @handle" → full pipeline.
Fallback: yt-dlp → FxTwitter → Playwright → Apify (paid backup).

What You Can Do

• "Scrape @garyvee" → Get all posts with real view counts, engagement rates
• "What's working for @hormozi?" → Top 30 posts by engagement, hook analysis
• "Show me all videos where Ryan is wearing glasses" → Deep visual analysis across scraped content
• "Compare @creator1 to @creator2" → Side-by-side performance, hooks, style DNA
• "Build a Notion database of @handle's content" → Full creator intelligence pipeline

The rule: Free first. Paid only when speed justifies cost. Instaloader → yt-dlp → FxTwitter → Playwright → Apify (last resort).

Level 2: Multi-Device Bridge (Optional)

For users with 2+ Macs. Turns separate machines into one unified Claude Code system — auto-syncing skills, hooks, scripts, references, and configs between devices. Zero manual commands. Zero maintenance.

What You Get

Prerequisites

• 2+ Macs on the same network (or reachable via hostname/IP)
• Remote Login enabled on each machine (System Settings → General → Sharing → Remote Login)
• Claude Code installed on each machine

Multi-Device Bridge

Machines on my network:

• Machine A: machine-a.local (primary)
• Machine B: machine-b.local (secondary)

SSH bridge is configured. "Do X on my [machine name]" = Claude uses SSH silently. Never show SSH commands. Never ask me to run terminal commands on the other machine. Just do it.

### Troubleshooting

| Problem | Fix |
|---------|-----|
| "Too many authentication failures" | Add `IdentitiesOnly yes` to `~/.ssh/config` |
| "Connection refused" | Enable Remote Login in System Settings |
| "Host key verification failed" | Run `ssh -o StrictHostKeyChecking=accept-new machine.local` once |
| Sync not running | Check: `launchctl list | grep claude` |
| Wrong hostname | Run `hostname` on target machine to verify |

### The 3 Levels

| Level | What You Get | Who It's For |
|-------|-------------|-------------|
| **Level 1: Solo Machine** | Hooks, CLAUDE.md, skills, MCP servers | Everyone starting out |
| **Level 2: Multi-Device Bridge** | + SSH bridge, auto-sync, natural language device control | Power users with 2+ Macs |
| **Level 3: Full Agent OS** | + 185 skills, Cognitive OS, clone system, automated backups to GitHub + Google Drive | Full-time builders |

---

*Built by Agent Luis. Updated automatically. Never outdated.*

## System Health Protocol — "Installed ≠ Working"

**The #1 mistake:** Installing a tool (MCP server, hook, memory layer) and assuming it works. Installing is 10%. WIRING it (telling Claude WHEN and HOW to call it) is 90%.

### The 3-Layer Health Check

| Layer | What to Check | How to Diagnose | Fix |
|-------|--------------|----------------|-----|
| **1. Installed?** | Is the MCP/hook/script in `settings.json` or `hooks/`? | `cat ~/.claude/settings.json \| grep "tool-name"` | Add it to settings.json |
| **2. Wired?** | Does CLAUDE.md tell Claude WHEN to call it? | Search CLAUDE.md for the tool name | Add usage rules to CLAUDE.md |
| **3. Firing?** | Has Claude actually CALLED it in a session? | Check enforcer tracking: `cat /tmp/claude-backup-track-*.json` | The enforcer nags until it fires |

### Common Failure Modes

| Symptom | Root Cause | Fix |
|---------|-----------|-----|
| "I installed 6 memory MCPs but none save" | Installed but not WIRED — CLAUDE.md has no rules for when to call them | Add 6-Layer Memory Stack table to CLAUDE.md |
| "Content ideas aren't being pushed" | Rule exists but says "after every build" — too vague | Make triggers explicit: "built a skill? Push idea NOW" |
| "Google Drive saves aren't happening" | Enforcer was matching `gdrive` but MCP tool is `google-drive` | Match EXACT MCP tool names in enforcer |
| "Hooks aren't firing" | Wrong event name or hook returns non-zero exit | Check `~/.claude/settings.json` hook config, test with `echo '{}' \| node hook.js` |
| "Supermemory runs out of credits" | Calling every session (30 credits/month) | Add daily cooldown (1x/day max) |
| "Session dies mid-build" | Context bloat — no size-gate or diet system | Install size-gate.js + skill-diet.py |
| "MCP tool not found" | Server crashed or not started | Restart Claude Code. Check `claude mcp list` |
| "Memory saves are slow / inconsistent / duplicating" | **Memory MCP bloat** — running 5+ memory servers simultaneously (mem0, memorix, omega-memory, memory-knowledge-graph, memory-service all active at once) | Pick ONE session memory (Engram) + ONE global recall (Supermemory). Remove the rest from your project config in `~/.claude.json`. More memory MCPs ≠ better memory. |
| "WhatsApp / TypeScript MCP won't start" | MCP config uses `node` to run a `.ts` file — Node can't execute TypeScript directly | Change command to `npx tsx` in `~/.claude.json`. Check if tsconfig has `noEmit: true` — if so, the project is tsx-native and was never meant to compile. |

### The Leila Audit Fix (Run This When Things Feel Broken)

```bash
# 1. Check what MCPs are installed
cat ~/.claude/settings.json | python3 -c "import sys,json; d=json.load(sys.stdin); print('\n'.join(d.get('mcpServers',{}).keys()))"

# 2. Check what hooks are installed
ls ~/.claude/hooks/

# 3. Check enforcer tracking (what's been called this session)
cat /tmp/claude-backup-track-*.json 2>/dev/null | python3 -m json.tool

# 4. Check if daily Leila Audit is running
launchctl list | grep leila

# 5. Check hook health (test each hook)
for f in ~/.claude/hooks/*.js; do echo "Testing $f..."; echo '{"tool_name":"test","tool_input":{},"session_id":"test"}' | node "$f" 2>&1 | head -1; done

The Golden Rule

Every tool needs 3 things to work:

1. Installation — the tool exists in settings.json
2. Wiring — CLAUDE.md has explicit rules for WHEN and HOW to call it
3. Enforcement — a hook or protocol that NAGS until it's done

If ANY of these 3 are missing, the tool is dead weight. The Leila Principle: "What's already in the building that nobody's using?"

Engineering Principles — DRY KISS YAGNI SOLID (Add to Your CLAUDE.md)

Every serious engineer follows these. Add to your CLAUDE.md:

## DRY KISS YAGNI SOLID (Engineering Laws)
- **DRY** — Don't Repeat Yourself. One source of truth. If it exists, import it.
- **KISS** — Keep It Simple. Simplest solution that works = best solution.
- **YAGNI** — You Ain't Gonna Need It. Build what's needed NOW, not "just in case."
- **SOLID** — Single Responsibility (1 skill = 1 job). Extend, don't rewrite.
Over-engineering check: "Is this the simplest version? Did someone ask for this? Does this already exist?"

Gate Protocol

• Pre-flight: Verify all dependencies and tools are available. Check file paths exist. Confirm no destructive operations without backup.
• Mid-flight: Log each operation. Verify intermediate outputs before proceeding to next step. Handle errors gracefully — never fail silently.
• Post-flight: Verify system state is correct. Confirm all outputs saved to expected locations. Test that the system works end-to-end.

Full Specification

Complete details, decision trees, protocols, and implementation specs: references/full-details.md

Change Log

• [2026-03-13]: Upgraded 3D Video Perception → 4D Human Perception (SEE+HEAR+FEEL+TIME-TRAVEL). Combined install steps 3+4. Added Social Media Intelligence section — full scraping stack (Instaloader, yt-dlp, FxTwitter, Playwright) with Python quick start + CLAUDE.md wire-up. Added Zero-Friction Install Gate — every skill ships with exact install commands.
• [2026-03-12]: Added Doctor Claude (/doctor-claude) — 5-question diagnostic → prescribe 1-3 skills. Anti-Dump Rule. Renamed from prescription-engine. Expanded Dave's Essential Apps → Dave's Full App Stack (35+ apps, 9 categories). Added skill versioning system (version: field in YAML frontmatter) + install-skill.sh one-command installer. Added skill tags system (tags: field in frontmatter).
• [2026-03-10]: Added System Health Protocol — "Installed ≠ Working" troubleshooting section. 3-layer health check, 7 common failure modes, Leila Audit fix commands, Golden Rule (Install + Wire + Enforce).
• [2026-03-10]: Added 6 new components: Daily Leila Audit, ROI Dashboard, Self-Improvement Protocol, 5 Core Clones, Dave's Essential Apps, Python Toolkit. ZIP at ~/Desktop/drey-starter-pack.zip.
• [2026-03-09]: Context Diet — split 627 lines into routing card (194 lines) + references/full-details.md (462 lines). Zero content deleted. By skill-diet.py.
• [2026-03-02]: War Room Tier 2 — content quality upgrade (gate→system).
• [2026-03-02]: War Room Tier 1 audit — skill structure verified and standardized.