How I Built an AI Agent That Automates My Entire Business with Claude Code

A deep dive into ATLAS — my personal AI automation system that handles emails, calendar, sales outreach, content creation, and client management. Built in 16 hours from Kampala, Uganda.

How I Built an AI Agent That Automates My Entire Business with Claude Code

Every morning at 4:00 AM, before I even touch my keyboard, my AI agent has already:

Read and categorised all my emails
Checked my calendar for today's meetings
Found 3 new potential clients
Drafted personalised outreach emails
Generated a full content package for LinkedIn, YouTube, and TikTok
Created AI-designed YouTube thumbnails
Set my top 3 priorities for the day

I call it ATLAS — Automated Tasks & Life Automation System. I built it in 16 hours using Claude Code, and it changed how I run my business.

Here is exactly how it works.

The Problem I Was Solving

I run Desishub Technologies from Kampala, Uganda. I build custom web apps, run 6 SaaS products, teach 300+ students, create YouTube content, and sell source code on Gumroad.

Before ATLAS, my typical day involved:

45 minutes reading and replying to emails
30 minutes researching potential clients
20 minutes staring at LinkedIn wondering what to post
90 minutes designing YouTube thumbnails in Canva
2 hours writing proposals from scratch

That is 4+ hours of admin. Every. Single. Day.

I was spending more time managing my business than building in it.

What ATLAS Actually Is

ATLAS is a collection of 28 slash commands and 5 AI skills that run inside Claude Code. It connects to my real tools via MCP (Model Context Protocol) servers:

Integration	What It Accesses
Gmail	Read emails, search, create drafts (never sends directly)
Google Calendar	Read events, check availability
YouTube Data API	Channel stats, competitor analysis, video engagement
Gemini AI	Generate YouTube thumbnails and images

The key safety rule: ATLAS never sends emails directly. It saves everything to Gmail Drafts with a [REVIEW NEEDED] prefix. I review and send manually.

The Architecture

Claude Code (Runtime)
    ├── .claude/commands/     (28 slash commands)
    ├── .claude/skills/       (5 AI skills)
    ├── .mcp.json            (MCP server configs)
    └── CLAUDE.md            (System instructions)

MCP Servers (Tool Layer)
    ├── google-workspace     (Gmail + Calendar)
    ├── youtube-data         (YouTube API)
    ├── github               (Repo management)
    └── nanobanana           (Gemini image generation)

Data Layer (D:\jb-brain\)
    ├── system/              (ICP, voice guide)
    ├── clients/active/      (Prospect files)
    ├── content/             (Daily packages + thumbnails)
    ├── logs/                (Morning briefs + EOD reviews)
    └── drafts/              (Email drafts)

Everything runs locally. No cloud service, no subscription. Just Claude Code and MCP servers.

The Commands That Matter Most

`/today` — Morning Brief

This is the first command I run every day. It:

Reads yesterday's end-of-day review
Checks Gmail for unread emails from the last 12 hours
Checks Google Calendar for today's events
Scans my client files for leads that need follow-up
Detects what type of content day it is (YouTube, TikTok, or Course day)
Compiles everything into a 250-word brief

Output looks like this:

🌅 GOOD MORNING JB — Sunday, April 5, 2026

🔄 YESTERDAY: 3 prospects emailed. Allan replied — wants Zoom call.
📌 TOP 3 TODAY:
1. Reply to Godwin on LinkedIn (messaged twice)
2. Book Allan's Zoom call (he's warm, don't let it cool)
3. Plan next week's YouTube videos

📅 1 event: WEBINAR ALERT at 3:30 PM
📧 Godwin messaged you again. Reply NOW.
💰 Quick win: Reply to Godwin. 2 minutes.

`/inbox` — Email Processing

This reads EVERY email, not just previews. For each email it:

Categorises: 🔴 Action Needed / 🟡 FYI / ⚪ Skip
Drafts a response for every action item
Saves all drafts to Gmail with [REVIEW NEEDED] prefix
Never sends anything directly

On Day 1, it processed 16 emails in 2 minutes. Manually, that would have taken 45 minutes.

`/find-prospects` — Sales Pipeline

Searches the web for 3 new prospects matching my ICP (Ideal Client Profile). For each one, it:

Researches their business (website, LinkedIn, press)
Checks them against my Anti-ICP filter (no know-it-alls, no low budgets)
Finds a specific gap in their business I can help with
Creates a prospect file in ~/jb-brain/clients/active/

`/thumbnail [title]` — AI Thumbnails

Generates 4 YouTube thumbnail concepts using Gemini AI in the style of Code With Antonio and JavaScript Mastery. Features:

3D device mockups (MacBook, browser windows, phones)
Glassmorphic icon badges
Bold typography with accent colors
Dark premium backgrounds
Auto-crops to 16:9

I went from spending 90 minutes per thumbnail in Canva to 2 minutes with /thumbnail.

Day 1 Results

Metric	Result
Emails processed	16
Prospects found	3
Outreach emails sent	4
Replies received	2 (Allan from ProGigFinder)
Warm leads	1 (Allan wants a Zoom call)
Content pieces generated	12+
Thumbnails created	16 (across 5 iterations)
Commands built	28
Time building ATLAS	16 hours
Time saved per day (estimated)	3-4 hours

The standout: Allan from ProGigFinder replied within 3 hours of receiving my cold email and asked for a Zoom call. The email was drafted by ATLAS, reviewed by me, and sent. The system works.

The Tech Stack

Runtime: Claude Code (Anthropic CLI)
AI Model: Claude Opus 4.6 (1M context)
Image Generation: Gemini 2.5 Flash Image API
Email: Google Workspace MCP
YouTube Data: YouTube Data API v3 via MCP
Data Storage: Markdown files in Obsidian (migrating to custom CRM)
Languages I build with: Go + Next.js (my production stack)

What I Learned

1. MCP servers are the real power. Claude Code without MCP is just a coding assistant. With MCP, it becomes a business operator. Connecting Gmail and Calendar was the inflection point.

2. Safety rules are non-negotiable. ATLAS never sends emails directly. Ever. Everything goes to drafts. This one rule prevents catastrophic mistakes.

3. Fewer commands, done well. I started with 22 commands and grew to 28. But /today, /inbox, /find-prospects, and /eod handle 80% of my daily workflow. Build the critical path first.

4. The prompt is everything. My YouTube thumbnail skill went through 5 complete rewrites before the output matched what I wanted. Iteration is the process.

5. This is not a toy project. Allan replied to my cold email in 3 hours. That is real revenue potential from an AI system I built in a day.

Source Code

The entire ATLAS system is open source:

GitHub: github.com/MUKE-coder/atlas

28 commands, 5 skills, MCP configurations, and the complete CLAUDE.md system instructions.

What Is Next

I am building a full CRM (ATLAS CRM Design Doc) with:

10 database modules (contacts, projects, courses, AI agents, content, revenue)
30+ MCP tools so Claude Code reads and writes directly to the database
Go API + Next.js admin dashboard
PostgreSQL + Redis

When complete, ATLAS will not log to markdown files — it will operate a full business database.

I am also offering AI agent setup as a service for professionals: doctors, lawyers, accountants, real estate agents. Same system, customised per career. UGX 1.5M-5M setup + monthly maintenance.

If you want your own ATLAS, reach out.