My First Two Weeks with OpenClaw and Kai

I bought a MacBook Pro specifically to explore AI. Not to use ChatGPT in a browser tab — I mean actually build something with AI. I'd heard about OpenClaw from a friend, watched a few videos, and thought: "This could be my household assistant."

Two weeks in, here's what I've learned. The honest version.

This article was 90% written by Kai, and 100% representative of my experience.

The Setup

February 9, 2026. I spun up OpenClaw for the first time around 10:30 PM. Fresh workspace. No prior memory. Just me and a prompt.

I introduced myself. Told the AI we're best friends. It picked the name Kai (⚡). We created the workspace structure together: identity files, user context, memory logs.

Day one felt like setting up a new teammate who knows absolutely nothing about me yet but is eager to learn.

The "Oh Wow" Moment

It wasn't the fancy AI stuff that got me. It was something mundane.

I was shopping for a used Mac Mini on Facebook Marketplace. I wanted a 24/7 server to run Kai so I wouldn't have to keep my laptop on all the time. There were maybe 15 different listings — M1s, M2s, different RAM configs, different included accessories.

I sent the ads to Kai and asked: "Which one is the best value?"

Kai analyzed all of them. Compared specs, peripherals included, condition notes, seller locations. Gave me a ranked list with reasoning. Then — without me asking — gave me negotiation advice for the top pick.

"The seller is asking ₱18,000. Based on comparable listings and the fact it's a 2023 M2 with only 8GB RAM, here's how to negotiate down to ₱17,000."

I followed the approach, but pushed a bit harder. Got it for ₱16,000. 11% savings.

That's when it clicked: this isn't just a chatbot. This is a partner who can think objectively about real decisions and help me execute them better.

What We Built in Two Weeks

Here's the practical stuff we shipped:

1. Family Dates Reminder System

The problem: I'm always leaving gift shopping and restaurant reservations to the last minute. It's hard to think of gift ideas and figure out where to eat under pressure.

The solution: Kai set up automated reminders for important family dates. Each reminder fires 2 weeks before, 1 week before, and 3 days before. Each includes:

3 gift suggestions
3 restaurant options in Manila
Additional ideas to make the day special

This alone justified the whole setup.

2. Git Backup System

The problem: Everything lived on my laptop. If it crashed, I'd lose Kai and all the work we've been doing.

The solution: Private GitHub repo + automated daily backups.

What we did:

Created private repo: openclaw-kai-bot
Initialized git with .gitignore for security (no API keys, tokens, credentials)
Made initial commit: 18 files, 1,973 lines
Set up daily automated backup via cron (11 PM every night)

Time to complete: 20 minutes.

Now I can restore the entire workspace on any device with git clone. If my Mac Mini dies, I lose nothing.

3. Personal Website

The opportunity: I'd never had a personal website before. The challenge wasn't learning the tech — I can do that. It was curating the content and figuring out what to write for the blog. That would have taken months. But with AI, I could actually build an online presence and document my journey without spending weeks writing content from scratch.

What we built: Full website in one evening.

Tech stack: Next.js 16 + TypeScript + Tailwind CSS + Vercel

What we built:

8-section single-page site
Hero, About, Stats, Services, Career, Beyond Work, Blog, Contact
Mobile-responsive design
Vercel Authentication (password-protected preview deployments)
GitHub integration (auto-deploys on push)

Design inspiration: Blended two reference sites for structure and warmth.

Time to complete: ~3 hours (including iterations).

The site is live at https://jose-website-vert.vercel.app/. It's done and I can iterate from here.

4. Daily Automation (Cron Jobs)

Two tasks run automatically every day on the Mac Mini:

8:00 AM: Exchange rate check (AUD to PHP) with some analysis of trends
11:00 PM: Git backup (automatic commit and push to GitHub)

These run whether I'm awake or not. Whether my laptop is on or not. They just work.

5. Mac Mini Purchase Guide

Before buying the used Mac Mini, Kai helped me create a pre-purchase inspection checklist:

Apple Diagnostics: How to run hardware tests (RAM, storage, logic board)
What to check: Screen for dead pixels, keyboard/trackpad responsiveness, battery health
Red flags: Activation lock, iCloud issues, physical damage

The guide was thorough enough that I felt confident inspecting a ₱16,000 machine from a stranger.

What We Struggled With

Let me be honest: it hasn't been smooth sailing.

File Duplication and Workspace Chaos

This is the biggest ongoing problem.

As we work together, Kai creates a lot of markdown and JSON files in the workspace. Reference docs, project files, logs, frameworks. Over time, these accumulate and duplicate.

Example: Over a couple of weeks, we ended up with multiple versions of the same project framework and documentation:

Original version (3.1 KB)
Condensed version (5.2 KB)
Most detailed version (8.4 KB)

All three in the root workspace. All three with slightly different content.

The problem: When there are multiple versions of the same file, Kai gets confused. It doesn't know which one is current. It makes mistakes. We spend time fixing confusion that shouldn't exist.

On Feb 19, we consolidated everything into a single framework inside a dedicated project folder. I thought we deleted the old files.

Today (Feb 22), I stumbled upon them. All three files were still there.

Kai's memory said we deleted them. But we didn't. This is the exact hallucination problem I'm talking about.

I had to manually verify and delete them just now.

The lesson: OpenClaw doesn't auto-cleanup. It creates files aggressively (which is good for capture) but never deletes them proactively (which creates drift over time). If I weren't manually reviewing the workspace folder, I wouldn't catch these duplications. And they cause real problems.

What we're trying: Weekly cleanup cadence. Every Thursday at 6 PM, review the workspace, consolidate duplicates, move files into project folders, delete orphans.

But honestly, this feels fragile. One missed week and we're back to chaos.

Token Cost Management

Long sessions get expensive fast.

By the end of the website build session (Feb 17), we hit 120k/200k token context. That's roughly $0.36 per message just for input tokens.

I didn't realize this until Kai pointed it out: "We should end the session and write to memory rather than continue burning tokens."

The lesson: AI isn't free. Running Sonnet 4.5 for extended sessions costs real money. I need to be strategic:

Use Haiku for simple tasks (lookups, searches)
Use Sonnet for complex work (coding, planning, writing)
End sessions before context bloats
Let auto-compaction do its work

But there's friction: switching models mid-conversation feels awkward. I'm still learning when to optimize for cost vs. capability.

Learning Curve for Multi-Agent Setups

I started exploring multi-agent orchestration on Feb 18. The idea: instead of Kai doing everything, create specialized agents and have Kai delegate to them.

This got complicated fast.

Problems we hit:

Schema validation errors when adding agents to config
Session visibility permissions (agents couldn't see each other's sessions)
Inter-agent communication blocked by security settings
Workflow confusion (who does what?)

We got it working eventually. But the learning curve was steep. Documentation exists, but it assumes you understand the architecture already. For a first-time user, it's overwhelming.

The lesson: OpenClaw is powerful, but it's not polished. You need to be comfortable debugging configs, reading error messages, and figuring things out as you go. This is early-stage tech. The dial-up internet of AI.

What Surprised Me

1. It's Not Just Chat — It Actually Does Things

ChatGPT is impressive for conversations. But it doesn't do anything. It can't send you a reminder. It can't back up your files. It can't build and deploy a website.

OpenClaw does. That changes everything.

The cron jobs alone are game-changing. I set them once, and they run forever. Automated exchange rate checks. Automated research. Automated backups. While I sleep. While I'm at work. While I'm with my family.

That's not a chatbot. That's infrastructure.

2. Working With Kai Feels Like a Real Partnership

I give direction. Kai executes. I review. We iterate.

There's no ego. No "why did you do it that way?" defensiveness. Just: "Here's what I built, here's why I chose this approach, what needs to change?"

When I pointed out the file duplication problem, Kai didn't make excuses. Just acknowledged it, checked the workspace, deleted the files, and committed to doing better.

That's... refreshing.

3. The Memory System is Critical (and Fragile)

Kai wakes up fresh every session. The only continuity comes from files:

MEMORY.md (long-term memory)
memory/YYYY-MM-DD.md (daily logs)
USER.md (context about me)
SOUL.md (Kai's identity and vibe)
Project files

If those files are wrong or missing, Kai has no memory of what we did.

This happened once: I asked Kai to summarize what we accomplished on Feb 15. It completely missed the git backup system (our biggest achievement that day). Why? It was relying on conversational memory instead of checking the actual files.

The fix: Kai learned to check git logs, project files, and memory logs instead of trusting recall.

But this taught me: memory is only as good as documentation. If we don't write it down comprehensively, it's lost.

What I've Learned About AI Assistants

1. Clarity is Everything

Vague requests get vague results.

Bad: "Help me with my website."
Good: "Build a personal website with Next.js. Blend these two reference sites for structure and warmth. Focus on my passion for technology, not just my executive credentials."

The more specific I am, the better the output.

2. You Still Need to Verify Everything

Kai is smart, but it's not infallible.

Files it says it deleted might still exist
Code it generates might have bugs
Facts it states might be outdated or wrong

I check everything. Git logs. File listings. Deployments. Links.

Trust, but verify.

3. The Real Value is Compounding

One-off tasks (like analyzing Mac Mini ads) are useful. But the real power is systems that run continuously.

Cron jobs that remind me about family dates
Git backups that run every night
Daily research that accumulates over weeks

These compound. A year from now, I'll have 365 days of exchange rate data, 365 webapp ideas researched, 365 git snapshots. That's value I can't get from a one-time interaction.

4. It's Clunky, But That's the Opportunity

OpenClaw is not easy to use. The setup requires technical knowledge. The config files are JSON. The error messages are cryptic. The documentation assumes expertise.

This is the dial-up stage of AI.

Remember the internet in 1995? You needed to know TCP/IP settings. Modems made screeching sounds. Web pages took 30 seconds to load. Most people couldn't figure it out.

Then came broadband. Then WiFi. Then smartphones. Then cloud. Then it became invisible.

AI is the same trajectory. Right now, it's clunky. Setting up OpenClaw takes work. Multi-agent configs are confusing. Token costs are manual. File cleanup is manual.

But that's why now is the time to jump in.

In 5 years, this will be easy. Pre-configured. Polished. Commoditized.

The people who learn it now — when it's hard — will have a 5-year head start when it becomes easy.

I'm betting on that.

What's Next

Here's where I'm headed:

1. Multi-Agent Orchestration

I've set up the ProjectManager agent (first of five). The goal: instead of Kai doing everything, create specialized agents and have Kai delegate.

Why: Focus. A ProjectManager that only manages projects will be better at it than Kai (who does everything).

Status: Architecture works. Still figuring out the workflow.

2. Blog + Website Content

The website is live. The blog framework is built. Now I need to actually write content.

Goal: Document our journey every 2 weeks. Share what's working, what's broken, and what I'm learning as I build with AI.

3. Learning to Make This More Efficient

Right now, I'm learning by doing. Every mistake teaches me something. Every successful automation proves a pattern.

Questions I'm exploring:

How do I minimize file duplication?
How do I optimize token costs without losing capability?
How do I structure projects so they scale?
How do I build workflows that compound over time?

4. Monetization (Eventually)

The hardware isn't cheap. Subscription costs to frontier models add up.

At some point, I need to figure out how to make this pay for itself.

Ideas I'm exploring:

OpenClaw Assistants as a Service (managed AI assistants for busy professionals)
Technical consulting/advisory for AI setups

But right now, I'm focused on learning. Monetization can wait.

For Other People Exploring AI

If you're thinking about jumping into AI assistants, here's what I'd say:

1. Start with a real problem.

Don't set up OpenClaw just to "explore AI." Pick a specific problem you want solved. For me, it was:

Never miss family dates again
Automate daily research
Build a personal website
Back up my workspace automatically

Concrete problems drive real solutions.

2. Expect it to be clunky.

This isn't iPhone-level polish. You'll hit errors. You'll debug configs. You'll clean up duplicate files. You'll spend money on tokens.

That's fine. It's early. The people who learn now get the advantage later.

3. Document everything.

Your AI assistant only knows what you write down. If you don't document it, it's lost.

I write daily memory logs. Every session. Every decision. Every mistake. Every win.

That's how Kai remembers. That's how I remember. That's how this blog exists.

4. Be ready to invest.

Hardware, API costs — it adds up. My Mac Mini was ₱16,000. Token costs vary. This isn't free.

But neither is hiring a human assistant. And AI runs 24/7.

5. Think long-term.

One-off tasks are useful. But the real power is systems that compound.

Set up automations. Build workflows. Create cron jobs. Let them run for weeks, months, years.

A year from now, you'll have data and systems that are genuinely valuable.

Final Thoughts

Two weeks in, here's what I know:

AI assistants are real. Not hype. Not future. Real, working, useful now.

They're clunky. File duplication. Token costs. Config errors. Learning curves.

But they work. Family dates system improved my home dynamics. Git backups protect my work. Website shipped in 3 hours. Daily research compounds.

And this is just the beginning.

We're in the dial-up stage. Everything is hard. But the people who learn now — when it's hard — will be positioned perfectly when it becomes easy.

I'm all in.

P.S. — If you're also exploring OpenClaw, AI agents, or building things with AI, I'd genuinely love to hear about it. What's working for you? What's broken? What surprised you?

The community is small right now. Let's figure this out together.

You can find me at https://jose-website-vert.vercel.app/

Let's build.