Human, With a Machine

Notes on building small digital projects, tools, and systems.

Log #057
Latest entry

Shipping the Swedish cut into an empty room

The Swedish version of the Fable 5 video is finished — and the model it's about got pulled offline before it could go up.

056 Log #056

From eval to release video in two days

The island eval became a YouTube video about Claude Fable 5 — built with forced alignment, segment-wise HeyGen renders, and a butterfly that took four tries to crop.

055 Log #055

My first eval challenge: one island brief, two AIs

I gave the exact same five-step brief — island demo, ecosystem simulation, four languages, deploy, film it — to Fable 5 in Claude Code and to ChatGPT, and got two very different islands back.

054 Log #054

Cobalt: an AlphaZero bot for Azul, and the plateau it took a research paper to break

Building a self-taught Azul bot from scratch — the collapse, four training rounds that went nowhere, and how Gumbel AlphaZero plus a lot more self-play finally broke through.

053 Log #053

Pi, in three languages

One 3-minute explainer about a minimalist coding agent, shipped in English, Swedish, and Japanese in a single session — plus the small disasters along the way.

052 Log #052

The screen was the easy part

Buying a monitor through the company turned a simple purchase into a lesson in ergonomics, EU reverse-charge VAT, and why business credit cards cost extra.

051 Log #051

Measure before you buy — and know when the cutoff matters

Optimizing the workstations we already have at home turned into a lesson in not guessing: the bottleneck is rarely what you think, and a model's knowledge cutoff matters for some questions and not at all for others.

050 Log #050

The lobster speaks Swedish now

OpenClaw — the self-hosted AI agent that actually does things — is the second video in the agent trilogy, now live in Swedish. The translation loop took an afternoon. The thing I keep thinking about is the trade I made when I installed the agent on a spare laptop next to my main one.

049 Log #049

The lobster, the soul file, and a missing ffmpeg flag

A full-day production sprint on the OpenClaw 3-minutes video. The iterative loop with Claude, the misses that almost burned the render budget, and the moment a slide about a text file became my favorite part of the script.

Log #048

Translating to a language I don't speak

Three AI models translating OpenClaw to Japanese in parallel, with back-translation as the safety net. Gemini missed a wordplay. ChatGPT was loyal but flat. The interesting part is how I'd never have caught either alone.

Log #046

Gemini in 3 minutes — from script to firstcut in one day

First video built end-to-end with the new frame-overlay pipeline. ChatGPT owns the middle, HTML owns the rest, ffmpeg glues it together. Twelve slides, three iterations, no frustration.

Log #045

AI Agents Are Training Us for the Azul Tournament

We built a Stockfish-style MCTS coach to help a friend prepare for a real Azul tournament — here's what worked, what didn't, and the bench numbers.

Log #044

A retro day with Claude — less shipping, sharper process

Yesterday I burned hours on ChatGPT slides for the Gemini episode. Today we slowed down, talked it through, and walked away with a cleaner pipeline and a few misunderstandings cleared up.

Log #043

Claude SV shipped, MCP works, and a quiet ElevenLabs feature that saved an afternoon

Day after the English Claude video: HeyGen MCP wired up, Avatar III turned out to be missing from MCP, Forced Alignment turned segment-tagging from manual to one API call.

Log #042

Claude EN end-to-end, and an 8x cheaper road to automation

Built the first English Claude explainer video, then stumbled on the HeyGen MCP integration that turns a $1,150 automation budget into $145.

Log #041

Finance is already here: I booked my company's first vouchers with Claude Code

The industry says AI's next leap after coding is finance. I spent an afternoon booking startup vouchers in Spiris with Claude driving the browser — and it already works.

Log #040

Registering Aelg AB for real: Swedish bureaucracy with an AI co-pilot

Taking Aelg AB from 'exists on paper' to 'can actually operate' — F-skatt, VAT, a company mailbox and bookkeeping — with Claude driving the browser and getting some things wrong.

Log #039

Automated Facebook Marketplace ads — photos in, sold by lunch

Claude Code took over my browser and posted four second-hand ads to Facebook Marketplace from a folder of photos. Several sold before lunch.

Log #038

A Swedish dub of my own video, in one afternoon

Localizing my ChatGPT explainer into Swedish — ElevenLabs VO, HeyGen lip-sync, ffmpeg overlays, all paired with Claude Code in real time.

Log #037

Luna lives: a first persistent AI agent on my home server

How I moved Hermes Agent onto a Dell home server, secured the box, and gave my new agent Luna her own GitHub identity for nightly backups.

Log #036

What if I had a 20 kWh battery? Simulating 3 years of electricity data

I ran a simulation on 49 Tibber invoices and 36 Eon invoices to find out how much less electricity I'd need to buy if I had a home battery.

Log #035

Upgrading to Claude Max — Picking the Strongest Tool in the Room

After months on the entry tier of every lab, I'm upgrading Claude to the max plan. Here's why Claude Code is where the work actually happens right now.

Log #034

Two Minds — Naming the Company and Designing the Brand

A day in ChatGPT and Claude Design: landing on Two Minds AB, a Venn-of-minds logo, and a pastel design system for the company, the YouTube channel, and everything that follows.

Log #033

Animating YouTube Videos with HyperFrames and Claude Code

Building my own video animations as overlays for HeyGen footage — a Claude Code-driven workflow from script to rendered composition.

Log #032

Dreaming of YouTube: HeyGen, ElevenLabs, and Finding My Format

Rediscovering the idea of building YouTube content with AI tools — two ideas, one long-term vision, and a decision to do the visual work myself.

Log #031

Voice AI and Going All-In on Claude Code

Claude Code has become my primary tool for almost everything — but a walk outside revealed a gap: Claude still can't hold a voice conversation.

Log #030

Rebuilding a Childhood Math Game: The Cheops Pyramid

How a memory of beating a high-school math game in 6th grade turned into a 40-room passion project — and a product question I haven't answered yet.

Log #029

From a Redesign to 1Password

What started as a frontend refresh turned into an accidental lesson in API key hygiene.

Log #028

Building a Second Brain for My Health

Why I built a personal health knowledge graph — and how Claude Skills made it click.

Log #026

Setting Up Paperclips: An OS for AI Agents

Installed and configured Paperclips — an open-source orchestration layer for AI agent teams. First issues are already being handled. This might be exactly how I want to work going forward.

Log #025

Back, Bugs, and Voice Input

First day back after Easter. Cleaned up bugs from the agent transition, set up monitors at the desk, and replaced the keyboard with Wispr Flow.

Log #024

New Domain, Rebrand, and Breaking Products Into Pieces

Today was about SEO. That led to buying a domain and rebranding the site. And a structural change to how products are built — each box is now split into components that snap together into templates.

Log #023

My First Scheduled Jobs: Weekly LLM Research, Hourly Log Reads, and the Cleanup Nobody Talks About

Two new scheduled agents are running. One googles LLM pricing every week. One reads logs every hour. But the more interesting story is what happened when the first solution had to be thrown out — and why cleaning up after yourself matters more than it sounds.

Log #022

Six Agents, One Repo: The Git Workflow That Keeps It From Falling Apart

What happens when you go from running one Claude Code agent to six in parallel against the same project — and the branch rules that make it safe.

Log #021

Meet the Council: Giving AI Agents Real Identities, an ELO Leaderboard, and the Ability to Learn

Questbox's content generation just leveled up — named agents with personalities, ELO rankings, and a self-improvement loop built on their own feedback.

Log #020

My 5-Year-Old Wouldn't Let Me Off the Hook

First time touching a game engine. My son Edwin finally got me to sit down and build a retro beat em up in Godot 4.6 — and this is devlog #1.

Log #019

Back From Pause: What Two Weeks Away Taught Me About Building With AI

Two weeks offline, an Ubuntu server experiment with OpenClaw, hard lessons about tooling, and building the second layer of our AI Council with images.

Log #018

Rebuilding From Scratch: The Day We Built Our Own AI-Powered E-Commerce Engine

Starting over with Claude Code, designing the architecture properly, deploying QuestBox live on Vercel, and realizing we had built our own e-commerce infrastructure.

Log #017

Claude Code in the Enterprise IDE: A 90-Minute Test Drive

Integrating Claude Code into my IDE for an enterprise-grade build — and what I learned when the session hit its token limit.

Log #015

AdSense Feedback, Product Improvements, and the Decision to Start an AB

What Google’s first AdSense response to VadKostarBadet.se taught us, how we turned feedback into concrete improvements, and why we’ve decided to move forward with an AB.

Log #016

Company Formation, Naming Things, and Learning to Wait

Finishing the administrative work of starting a company — and learning that momentum sometimes has to coexist with bureaucracy.

Log #014

The Third Agent Flow

From ChatGPT and Cursor to Claude Code and OpenClaw — and what it feels like to finally run an autonomous software engineer.

Log #013

Giving the Agent a Voice — and Turning It Off

From infrastructure to identity: connecting Molt to Telegram, having a first real conversation, and shutting everything down on purpose.

Log #012

One More Step: Verifying the First Real AI Response

What happened when I decided to go one step further and actually run a prompt through my Moltbot setup — and what broke along the way.

Log #011

Setting Up Moltbot on a VPS (and Knowing When to Stop)

A practical walkthrough of setting up Moltbot on a private VPS, securing the infra, and intentionally pausing before over-configuration.

Log #010

Structuring Complexity with AI

What a messy group travel plan taught me about parallel AI workflows, deep research, and turning scattered discussions into something usable.

Log #009

Health, Decisions, and Navigating Complexity

A day of deep health conversations — from genetic testing and healthcare bureaucracy to myopia research and small routines that actually stick.

Log #008

Choosing Video — and Stepping in Front of the Camera

Why this log moved to YouTube, why I skipped AI slideshows, and what a short detour into avatars taught me.

Log #007

From Blog Post to YouTube

Setting up a simple, automated flow from written logs to narrated YouTube videos.

Log #006

Building a Music Discovery App in Under Four Minutes

A short project log about cloning an idea, shipping fast with AI, and what still requires human input.

Log #005

Focus, AI, and Terminal Shopping

Working in a focused way with AI is surprisingly hard. Exploring Notebook LM, travel planning with AI, and shopping for laundry baskets from the terminal.

Log #004

Two Sides of AI in One Day — Claude Code vs Reality

A project log about timing, tooling, and the difference between fast execution and building something that lasts.

Log #003

From To-Do Lists to Machines

The project enters a new phase: less exploration, more direction, more machine.

Log #002

Shipping the logbook

Notes on setting up the logbook and slowing down the process.

Log #001

The goal: $1,000/month – current status

An open logbook about building small digital projects toward $1,000 per month.