I write because I might be a few steps ahead of someone who wants to learn something new. Someone who needs a path forward.

We all have a unique mix of skills, behaviors and experiences. This combination becomes your unfair advantage when you share it.

Share your experiences.

“I might be wrong” is a cornerstone of how I manage. It keeps me flexible and open to reconsidering when someone explains I’m off base.

I’ve been taking improv classes for a couple of months now.

The thing I love most about improvisation is the feeling of a team working together toward a goal.

When we have aligned goals and clear structure, we trust each other and we want each other to succeed.

Then it is a marvel.

The problem isn’t with AI’s capabilities.

Companies are trying to automate chaos and expecting AI to magically create order from disorder.

I’ve been thinking about this lately. Not whether confirmation bias exists (it always has) but whether we’ve built an entire infrastructure around it.

The algorithms show us what we already believe (the filter bubble). ChatGPT is quick to tell us we’re right about things. And we’ve developed a whole subculture of just… cutting people off when they disagree. Unfriend, unfollow, block.

It’s not that any of these things are necessarily wrong on their own. But together they create something like a confirmation bias feedback loop. You think something, the algorithm reinforces it, the AI validates it, and anyone who might challenge it gets removed.

Maybe every age has its dominant cognitive trap. Ours just happens to come with decent UX.

Claude Code Web gave everyone $1000 credit to experiment with. I decided to build a restaurant catalogue chatbot for Umeå with full menu data.

The stack is Cloudflare: D1 for the database, Vectorize for semantic search, R2 for storage. Gemini parses uploaded menus. Perplexity fills in restaurant information. Sounds complicated but it’s fairly straightforward.

The workflow is what makes this interesting. Claude Code Web connects to a GitHub repo. I ask for a feature. It reads the code, makes edits, creates a PR. Cloudflare automatically deploys a feature branch with its own URL.

I test it on that URL. Ask for fixes if needed. Approve. It merges to main. Production deploys.

All of this from my phone.

I’ve made around 90 PRs. Spent about $100. That’s roughly $1.11 per shipped feature.

The constraint is being clear upfront. If I’m vague, Claude goes in the wrong direction and I waste iterations. When I’m precise, it usually nails the implementation.

The only real limitation is D1 migrations. Can’t run those from mobile. Though I could probably solve that with a GitHub Action.

I also bought a domain for 9 kronor.

It’s been a fun weekend experiment.

For the longest time there was a disagreement. It was about tabs vs spaces. Developers argued in threads and pull requests. No conclusion to what was better.

I jokingly always said ”one tab and one space” and very few laughed. That is how serious the ”war” was.

Then code completion, prettier and lint came along and the discussion ended.

I wonder what discussions we have today that will feel ancient in a couple of years time.

Micro-services? AI assisted coding? Type in JavaScript? The cloud?

Until it becomes a habit, breaking a routine requires force.

With impostor syndrome, I see the complete me. What I’m good at and what I’m not good at.

When other people look at people with impostor syndrome, they tend to see what those people are good at.

That’s the asymmetry.

I experience every moment of struggle, every gap in knowledge, every time I’m winging it.

Other people only see the output, the moments where I show up and deliver.

They don’t see the internal uncertainty or the things I avoid because I know I can’t do them well.

So they build an image of me from the highlights while I’m working with the full unedited footage.​​​​​​​​​​​​​​​​

They see the Instagram feed. I see all the photos that didn’t make the cut.

I’ve been thinking about transformation a lot recently because… well… AI and while I’ve used it plenty for myself I’ve been struggling with the transformation of a whole organisation.

And I think this was my main blocker, starting to big. With this in mind I started dabbling with alternatives and created this acronym REF.

Result. Eliminate. Focus.

Most AI transformation advice tells you to think big. Build a comprehensive strategy. Map every department. Create governance frameworks. Six month rollouts.

This is why nothing happens.

REF is the transformational version of MVP.

Start with one workflow where you can show a result this week. Not next quarter. This week. Someone writes an email in 30 seconds instead of 30 minutes. That’s the result that changes belief.

Then eliminate whatever makes it hard to do again. Is Claude blocked by IT? Fix that. Do people not know it exists? Tell them. Is the prompt buried in a shared doc? Make it a bookmark. Eliminate friction until using AI is easier than not using it.

And focus means picking one thing. Not “transform customer service and sales and operations and HR.” Pick customer service. Or pick one team in customer service. Or pick one type of email that one team sends.

The paradox is that narrow actually spreads faster than broad. When one team gets a 10x improvement on something specific, other teams notice. They ask how. They want in.

Trying to transform everything at once means no one experiences anything changing.