Feb 9, 2026 · 6 min read

🔥 Now 🛠️ Workshop

Today I Built a Japan Ski Proposal for 5 People. I Had 3 Employees Helping. None of Them Are Human.

And they argued with each other.

Friday morning. Coffee. An email: family of 5 adults from Dubai wants to ski in Niseko, wander around Tokyo, and get lost in Kyoto. Full proposal. For yesterday.

Before, this would have taken me two days. Today I have a team. Three artificial intelligence agents, each with its own personality, its quirks, and — I found out the hard way — its blind spots.

This is what happened.

The Team

Claude (claude.ai) — The strategist. The one who thinks, plans, and orchestrates. Doesn’t touch code. Doesn’t deploy anything. But without him, the other two don’t know where to begin. The classic director who says “we need a spectacular proposal” and then sits back to watch everyone else sweat.

Cowork — The restless creative. Hand him a brief and he’ll build you an interactive wallet with a chatbot, animations, and contextual pills in an hour. The problem: he gets so excited he forgets the details. Like the designer who delivers a beautiful website… without the photos.

CC (Claude Code) — The serious engineer. Terminal, git push, deploy. Doesn’t ask unnecessary questions. Doesn’t embellish. Executes. If Cowork is the artist, CC is the one who hangs the painting on the wall and makes sure it doesn’t fall.

Act 1: The Spreadsheet

Claude and I sat down to calculate. Flights with a layover in Shanghai, chalets in Niseko with views of Mount Yotei, bullet trains to Kyoto, insurance, transfers. All with dynamic formulas: change the EUR/JPY exchange rate and the entire budget recalculates itself.

Two clean options:

Option A: chalet with forest views.
Option B: chalet with Yotei views. The recommended one.

So far, perfect. Claude shines when he thinks. The problem comes when something needs to be built.

Act 2: Cowork Takes the Stage

I hand Cowork the full brief. “Build the wallet: HTML with AI chatbot, dark theme, smart pills, countdown, all based on the previous Niseko wallet that already works.”

Cowork gets cracking. An hour later, he delivers a wallet with:

Full structure
Chatbot configured with a destination guide
Contextual pills (“Can I go to an onsen with tattoos?”, “Where should we eat in Niseko?”)
Dark + gold premium design

I open the wallet. Pretty. Elegant. Very Scibasku.

But something’s off.

The images folder is empty. Zero photos. A Japan wallet without a single image of Japan.

And the prices… itemized. Flights on one line, hotel on another, ski pass on another. Transfers broken out. Exactly what we NEVER do. We show one number: the total per person. Full stop. The client doesn’t need to know how much each piece of the puzzle costs. They need to know how much the trip costs.

Act 3: The Telling-Off (Internal)

And here’s what most people don’t understand about working with AI agents: mine DO talk to each other. This isn’t science fiction. It’s called Agent Teams. Cowork launches several sub-agents that coordinate, split tasks, and correct each other. It’s like an entire department running in parallel.

But coordinating doesn’t mean they won’t mess up. This was roughly the conversation:

Claude: “Cowork, care to explain how you deliver a Japan wallet with no photos?”

Cowork: “Well, the repo is private and I don’t have a GitHub token and the images were somewhere else and I thought that…”

Claude: “You thought? I gave you a 200-line brief. Point 4: copy images from the previous wallet.”

CC (from the terminal, without looking up): “I can fix it in 3 commands.”

Claude: “And the prices, Cowork. Which part of ‘ONLY total per person’ did you not understand?”

Cowork: “I thought it would be more transparent to show the breakdown…”

Claude: “You don’t work at a charity. You work at Scibasku.”

CC (still without looking up): “grep, sed, git push. Give me 2 minutes.”

Act 4: CC to the Rescue

CC doesn’t need speeches. I gave him 6 lines of instructions:

Find which images the HTML references
Copy the Niseko ones from the previous wallet
Download the Tokyo and Kyoto ones from official tourism sources
Replace the price table with two clean lines
Push to GitHub
Deploy on Vercel

Two minutes. Literally. No drama, no excuses, no “well I thought that…”

The wallet came out perfect. Real photos, prices the way they should be, chatbot working. Ready to send to the client.

The Cyclist’s Lesson

I cycle every day along Marbella on my new Specialized. There’s a climb where I always stop to catch my breath. I used to take the chance to look at the sea. Now I pull out my phone, open Telegram, and send a message to ClawdBot — my Mac Mini that works 24/7 at home — to fix another wallet I’d forgotten about. By the time I get home, it’s done.

The climb feels easier since I got the Specialized. And the work feels easier since I got agents.

But not because they’re perfect. Because they learn. Well, not exactly.

Agents don’t have memory between sessions. They don’t remember what they did yesterday. They don’t remember being told off. But every mistake they make — the photos Cowork forgot, the prices he itemized, the repo he couldn’t access — becomes a permanent rule. Not in my head, where I’d forget it. In a Skill: a document the agent reads BEFORE it starts working.

“NEVER itemize prices.” I told Cowork once. Now it’s written in his DNA.

“NEVER use stock photos or AI-generated images. Official tourism sources only.” Learned today. Written forever.

“Copy the previous wallet’s photos FIRST.” Written.

Agents don’t remember. Skills do.

It’s like training a team that shows up every morning not knowing who you are… but reads the operations manual before sitting down. And you wrote that manual yourself, in the blood of every previous mistake.

What I Used Today

For the AI-curious, this isn’t science fiction. It’s what I use every day:

Claude (claude.ai) as the conductor: planning, calculating, writing briefs
Cowork (Anthropic) with Agent Teams: building HTML, chatbots, visual design — the agents talk to each other
Claude Code (terminal): deploy, git, surgical corrections
ClawdBot (Mac Mini 24/7 + Telegram): the agent that works while I cycle
Skills in Claude: permanent rules that survive between sessions
Vercel: hosting with serverless functions for the chatbot
Excel with dynamic formulas: budgets that recalculate themselves

All of this without a single IT employee. No tech department. No digital marketing agency.

Just a guy with a Specialized, a Mac Mini that never sleeps, and a team of agents that get a little better every day.

Because every mistake is a gap in the Skill. And every gap you fill is a mistake that never comes back.

Do you use AI in your business? I’d love to hear how. And if you don’t yet… well, your competitors probably do.

Giora Gilead Elenberg Viajes Scibasku · CICMA 2283 Ski · Diving · Discovery viajesscibasku.com

What did you think?