Long Journey House // SF // March 2026

Build Your
AI Agent

How I built an AI chief of staff without writing code.

Hey, I'm Timour.
I'm the founder of Edge City. My job involves a lot of different contexts - partnerships, comms, strategy, team coordination, content, programming, fundraising - all happening across a dozen different tools at once. Over the last few months I built a personal AI agent that handles about 30% of my job. I'm not an engineer - I don't write code. Tonight I'm going to walk you through how it works and how you can build your own.

This page is both the session guide and the resource you take home. Everything is clickable. Sections with a ▼ toggle expand for more detail.

But first - let's calibrate. Hands up: who uses ChatGPT or Claude regularly? Who's tried Claude Desktop? Who's heard of MCP or agents? What's the most interesting thing you've done with AI?
Scan to follow along on your phone

build-your-agent-sf.vercel.app

Quick Start - Three Levels
Level 1: Get started (5 min)
  • Go to claude.ai and get a subscription ($20/mo Pro or $100/mo Max)
  • Download Claude Desktop
  • Open Cowork, set up a Project with instructions about who you are and what you do
  • That's it. You now have a personalized AI assistant. Everything else builds from here.
Level 2: Claude Code (15 min)
  • Option A - Cursor (friendlier): Download Cursor, go to Extensions, search "Claude Code for VS Code", install it. Open it with CMD+SHIFT+P > "Claude Code: Open in New Tab"
  • Option B - Terminal (more powerful): Open Terminal (Mac: CMD+Space > "Terminal"). Run: npm install -g @anthropic-ai/claude-code then claude
  • Create a CLAUDE.md file: ask Claude "Help me create an instruction file for my work"
  • You now have the engine that powers R2. Same brain, but it can build things.
Level 3: Full AI Chief of Staff (30 min)
  • With Claude Code running, install CoStack:
  • git clone https://github.com/timkosters/costack.git ~/.claude/skills/costack && cd ~/.claude/skills/costack && ./setup.sh
  • Type /bootstrap - CoStack interviews you for 20 minutes about your work, team, and tools
  • You now have 9 skills, a context map, and a self-improving system. Welcome to the club.
An AI Agent Has Four Parts
A brain (Claude), memory (a notes folder), instructions (a text file about you), and connections to your tools. The AI is already smart. What it's missing is your context.
YOU (prompts + decisions) BRAIN Claude Opus 4.6 reads / writes / reasons / builds INSTRUCTIONS CLAUDE.md "who you are, how you work, what to do and not do" MEMORY Obsidian + QMD 5,280 files indexed searchable by meaning CONNECTIONS MCP (12+ tools) Gmail, Calendar, Notion Telegram, CRM, more SKILLS /context-update /context-collect /derive /drift /content-sweep /humanize /reflect reusable commands that orchestrate brain + memory + connections
What R2 (my agent) actually does
  • Morning scan: Reads my email, Telegram groups, and meeting notes. Summarizes what changed overnight. Catches ~95% of context.
  • Call prep: I type /context-collect Jennifer and it pulls everything about her from every source into one page.
  • Drafts everything: Follow-up emails, partnership proposals, team updates - with full context, not generic AI slop.
  • Tracks commitments: Surfaces things I said I'd do but haven't, across 15+ active projects.
  • Self-improves: Compares my drafts to what I actually sent and learns the differences. When something goes wrong, I run /reflect - it traces the problem and updates its own instructions. Over time the system genuinely gets smarter.

This used to take 2-3 hours a day. Now it takes 30 minutes.

The self-improvement loop
  • /reflect - Analyzes patterns across sessions, proposes improvements to its own instructions
  • /diary - End-of-session capture: what worked, what didn't, what surprised me
  • Draft review - When I publish something R2 drafted, it compares draft vs. what I actually sent. Logs the differences. "Timour softened the tone and added a personal detail." Over time, drafts get closer to how I actually write.
  • Weekend research - R2 searches Twitter, GitHub, and the web for new tools or approaches that could improve our setup. Usually finds a dozen things; we pick 1-2 to implement.
Three products, one brain
  • Claude Chat (browser) - you ask, it answers. Like texting a smart friend. Good starting point.
  • Claude Cowork (desktop app) - you describe the outcome, it executes. Reads your files, accesses your calendar. Like delegating to an employee. Start here.
  • Claude Code (terminal or Cursor) - the power tool. Builds and deploys things. This is what R2 runs on. You can run it in the terminal directly, or inside Cursor (a code editor with a nice UI - don't let the word "code editor" scare you, you won't write code).
The Building Blocks
📋
CLAUDE.md - Instructions
A text file telling the AI who you are and how you work. Claude reads it every session. One good instruction file changes everything.
🔌
MCP - Tool Connections
Open standard for connecting AI to Gmail, Calendar, Notion, Telegram, your CRM. Hundreds of connectors. Start with email + calendar.
📂
Obsidian - Memory
Note-taking app = just markdown files on your computer. Claude reads/writes naturally. Wiki-links create a knowledge graph. The agent's long-term memory.
🔍
QMD - Semantic Search
By Tobi Lutke (Shopify CEO). Indexes your markdown files and searches by meaning, not keywords. 5,280 files indexed in my setup.
What I have connected (12+ sources)
  • Gmail, Google Calendar, Google Drive
  • Telegram (custom scripts)
  • iMessage (reads local database on Mac)
  • Notion (team workspace)
  • Granola (meeting transcripts - probably my most useful connection)
  • Attio CRM
  • Todoist (tasks)
  • Beeper (WhatsApp, Signal)
  • Typefully (social publishing)
  • PostgreSQL database (ticketing system)
  • QMD (semantic search across 5,280 files)
Skills
A skill is a reusable command written in plain English. Type /skill-name and Claude follows the recipe. You write it once, run it whenever.
SkillWhat it does
/context-updateMorning scan. Sweeps all sources, builds a map of my world.
/context-collect [name]Everything about one person from every source. Game-changer for call prep.
/deriveReasons about what the context means. Spots gaps, flags deadlines.
/driftCompares what I said I'd do vs what I actually did.
/content-sweepScans Twitter, articles, notes for content ideas. Drafts posts.
/humanizeStrips AI patterns from writing. 24-pattern checklist.
/deploy-siteBuilds and deploys a site. How this page was made.
/reflectAnalyzes patterns. System gets smarter over time.
Pro tip: Feed things from Twitter to your agent

Find an interesting article or thread on X. Send the link to your agent. If it says "I can't access Twitter," just ask it to sort that out - it can install the tools it needs. This is how I discover and integrate new ideas constantly.

How to make your own skill
  • Create a folder with a SKILL.md file inside
  • Write the instructions in plain English
  • Drop it in your skills folder
  • Done. Or ask Claude: "Help me create a skill that scans my email every morning."
Two Tips That'll Save You Grief
Plan mode (Shift+Tab)

In the Claude Code text box, use Shift+Tab to cycle through modes. Select Plan mode for big tasks. Instead of diving straight in, Claude thinks through the approach, shows you the plan, and you approve before it executes. Way less anxiety than watching it freestyle. Your best friend early on.

Bypass permissions (once you trust it)

By default, Claude asks for approval at every step. Good for building trust. Once you're comfortable, Shift+Tab to turn on bypass mode. Claude just runs. You're still watching, you can stop anytime. I wouldn't do this on day one, but it's how I work most of the time now.

Three Principles
1. Ask Claude about Claude

Don't Google "how to use Claude." Ask Claude itself. It knows its own capabilities, limitations, and can walk you through any setup step by step. This is the single most underused trick.

2. Context is the whole game

The AI is already smart. What it's missing is YOUR context. Your role, your projects, your preferences. One good instruction file changes everything.

3. Set boundaries early

Agent-first work has the same dopamine mechanics as a video game. Time-box your tinkering. Keep a "what shipped" log. The guardrails should come before they feel necessary.

Things I Built
I'm not an engineer. Claude writes all the code. I describe what I want and we iterate.
My workflow for building custom sites
  • Go to variant.com (AI design tool) and prompt a design vibe
  • Copy the generated code
  • Paste it into Claude Code: "Use this design to make a site with this content: [your content]"
  • Iterate on the design together
  • Deploy with /deploy-site - live URL in seconds

This is how this page was built. The whole workflow takes about an hour.

🎬
Video Semantic Search
I saw Google released a new embeddings model. Told my agent to process 1TB of video content. Now we have semantic search across 5,500 clips. "sunset ceremony" finds the right footage instantly.
🌌
Reading Cosmos
3D visualization of 8 years of reading. 329 books, 1,320 highlights, semantic connections.
🏛
Partnership Proposal (demo)
A partnership proposal as a microsite instead of a PDF. Built and deployed in ~2 hours.
🤖
Edge Esmeralda (built by Esme)
My Telegram agent built and published this herself. An agent pitching humans to come to our event.
📦
CoStack (Open Source)
The system behind R2, open-sourced. 9 skills. Run /bootstrap to set up your own.
🎨
Personal Site Reskin (live demo)
My personal site content in a TouchDesigner-inspired design. Built tonight to show the variant → Claude Code → deploy workflow.
Your Setup Checklist
Tap each item as you complete it. This saves locally in your browser.
  • Level 1: Get a Claude subscription at claude.ai
  • Level 1: Download Claude Desktop
  • Level 1: Set up a Project in Cowork with instructions about your work
  • Level 2: Install Claude Code (Cursor or Terminal)
  • Level 2: Create a CLAUDE.md instruction file
  • Level 2: Set up a notes folder (Obsidian or plain folder)
  • Level 3: Clone CoStack and run /bootstrap
Get the Tools
Everything you need, in order. See the Quick Start at the top for which level to aim for.
Claude Desktop
Level 1. The app - Chat, Cowork, and Code. $20/mo Pro, $100/mo Max.
claude.ai/download
Cursor
Level 2. Code editor with a friendly UI. Install the "Claude Code for VS Code" extension inside it. Don't let "code editor" scare you.
cursor.com
>_
Claude Code (Terminal)
Level 2. Run npm install -g @anthropic-ai/claude-code then claude. More powerful, same brain.
📦
CoStack
Level 3. The open-source AI chief of staff. Clone, run /bootstrap, 20 min interview, done.
github.com/timkosters/costack
Essential Tools
📒
Obsidian
Free. Your agent's memory. Markdown files + wiki-links.
🎙
Wispr Flow
Voice-to-text everywhere. Can't go back once you start.
🔊
Monologue
Voice transcription that learns your writing style and adapts to match.
Handy Computer
AI-powered voice assistant for your Mac. Natural conversation with your computer.
📝
Granola
AI meeting notes. No bot joins your call. Structured notes with action items.
Cursor
AI code editor. Vibe coding + Claude Code in a visual environment.
🎮
omar.build
Interactive 3D web experiences. Copy the code, paste to your agent, remix into anything.
🎨
Variant
Prompt any UI design, get code. Steal designs for personal sites, pitch decks, anything.
Tools People Shared
Cool things that came up during discussion.
🛡
Hermes
Agent platform with strong security, great onboarding, and self-learning memory. Multiple people recommended it.
🧠
Letta
Memory-first agent platform. Desktop app + coding agent. The agent manages its own memory as you go.
🌐
Index Network
Finds connections from your data exhaust (Twitter, Substack). Privacy-preserving with zero-knowledge cryptography.
🌌
Nebula
Agent platform from the Founders Inc team. Strong narrative and growing fast.
OpenClaw
Open-source system for a persistent AI agent on Telegram, WhatsApp, or Discord. My agent Esme runs on it ($12/month).
🤖
OpenClaw
Your AI that's always on. Message it from anywhere.
Power move: Use Claude Code to control your OpenClaw

Your Claude Code agent can SSH into your OpenClaw server and manage it directly. It can also send messages to your OpenClaw agent via Telegram. So R2 (my local agent) can set up, configure, and communicate with Esme (my server agent). One agent managing another.

Easy auth: Use your OpenAI subscription

OpenClaw supports OAuth with your existing OpenAI subscription - no separate API key management needed. Makes getting started much simpler.

Honest take on OpenClaw
  • Powerful but real tradeoffs
  • Pay-per-use API costs (some people hit $200+/month)
  • Always-on = you need trust + guardrails
  • Review community skills before installing them
  • For daily work, I prefer Claude Code on a flat $100/month
  • OpenClaw is great for bots and persistent agents
Background Reading
📖
How to Build Your Own AI Agent
Full R2 setup walkthrough. Written for non-coders.
📓
Two Weeks of Agent-First Work
The gains, the dopamine trap, and the guardrails.
Keep Up
🌐
Edge City
Pop-up cities for builders. Edge Esmeralda 2026: May 30 - June 27, Healdsburg CA.
edgecity.live
Seq. Run: 7F3A Build Your AI Agent // Long Journey House SF