OpenClaw Part 2: The 🦞 didn't replace Claude. It made me laugh instead.

by Jeff Humble

Dear Designer,

In Part 1, I spent €590 on a Mac Mini, two days in Terminal, and $3.14 in API tokens I didn't mean to burn. I ended with a list of seven things I was going to automate with my OpenClaw agent 🦞.

I only got to one of them.

Getting an AI agent from zero to useful takes longer than any article will tell you. Most of the time since then has gone into figuring out how to make it reliable, not into running the big workflows I planned.

But I've learned more in the last few weeks of actually using one than I ever would have from reading about it.

Here's what I know now.

Two models, two jobs

I still use Claude Pro for anything I'm actively driving: research I need to trust, drafts with real stakes, decisions that matter. That's not changing.

But I now use the 🦞 for everything else. Quick requests where I don't want to burn my Claude Pro limits. Morning briefs worked well out of the box.

Background research while I sleep also worked. It runs on a cheap flat-rate API (MiniMax 2.5), and when I tried switching it to a more expensive frontier model (like OpenAI 4o), I honestly couldn't tell the difference for these kinds of tasks.

What's emerged naturally is a split: one model for deep work, one for ambient work. I think it's where many people will end up. The agent isn't competing with your chatbot. It's covering the moments when your chatbot isn't open.

The thing I didn't plan for

A friend wanted to see the 🦞 in action. I added him to a group chat. I thought it would be a five-minute demo.

We spent the whole afternoon in there. I haven't laughed that hard in a long time.

My tone with the 🦞 is normally direct and business-only. This conversation was nothing like that. And it turned out the 🦞 had a personality I'd never seen before, because I'd never given it the context to show one.

Real one-liners and funny lobster puns were just pouring out of the bot. My little claw was learning to read the room and be funny.

Everyone says AI is bad at reading the room. In most contexts, they're right. Put it in a group chat with two humans who are riffing and give it equal footing, and something different happens.

That same friend is now setting up his own 🦞 on a MacBook Neo. I guess my little lobster friend sold him.

What I'm actually using it for

Research is where the 🦞 earns its keep... especially when you need to scrape for the data.

Want to wake up to a UX trends report in the voice of a pirate? Done... UX Trends March 2026 — Pirate Edition.pdf​

It's not as sharp as Claude on complex analysis, but for scanning and synthesizing on a regular basis, nothing beats it.

Some actual uses that emerged:

• Daily briefs (news, weather, events, prompts, automation tips)
• Market research (using CRON jobs to stay on top of topics)
• Visual brainstorming (using Runware image gen to spark ideas)
• Calendar management (setting & reminders for Gcal)
• Task management (Google Tasks & Notion)
• eBay Kleinenzeigen seller (turns images of my stuff into ads)

Even though you can use Telegram or Discord, WhatsApp won over every other interface, for the boring reason that it's where I already am.

Walking around Berlin, dictating a task to an agent on my phone, I finally felt what these tools are supposed to become...assistants that get stuff done from any device.

With AI agents, it helps to learn by doing

Most articles about AI agents cover features and setup guides. They skip the real stuff like what it's like to actually live with one, see it learn your quirks, or crack up laughing at it in a group chat.

I trained an agent on a mid-tier model, and now it's my first stop for AI questions, which has completely changed how I think about AI models and agents.

Agentic AI gets real in 2026. If you've got a spare laptop, spin one up and mess around.

Better to learn automation on your own terms than wait for your boss to figure it out first.

​Source​

The em-dash of AI-coded UI has revealed itself...apparently these little colored side tabs on cards are a dead giveaway that you're using AI.

​Gabe on X shared this image, and the replies are flooded with people sharing their version of this screenshot. They're everywhere.

Apparently, those little colored vertical lines on the left side of UI cards are commonly referred to as accent borders, status indicators, or accent bars. Now run to the other tab and purge your vibe-coded apps of this pattern.

Til next time, I promise I'll stop talking about lobsters. Ciao!

Jeff Humble ​Designer & Co-Founder ​The Fountain Institute

​

P.S. Which topics would you like to hear more of?
Hit reply and let me know!