Putting AI to work

I've written before that I built my own CRM, and that I did it eleven times faster. It has grown since then. It now automatically logs who I've written to from the moment I send an email. I can register a phone call with five words in the terminal. It creates a daily summary that distinguishes between a meeting, a call, and an email, and tells me about the week in prose. It even gives me a little streak counter and some experience points, so I remember to maintain the filing system.

I recently made a small contribution to Nushell — the open source shell I use daily. It's a tiny improvement in how custom completers work, but it solves a specific friction I encountered every single time I had to use my own terminal tool. Here's what it's about.

I've built a small Nushell module called yolay.nu. It's about 470 lines of code. It's the most productivity-enhancing thing I've written this year, and I think it points to something broader about how AI should be integrated into a workday—not just mine.

There are two stories about AI and code. One is the one everyone talks about: that language models write programs for us. The other is the one that is more important, but that almost no one talks about: that code keeps language models in check.

The two stories are the same story. But you have to tell them together to understand where the field is actually heading.

For a year, I've been writing notes in a system called IWE. It's a small command-line tool that treats markdown files as a graph — notes link to each other, and you can navigate, search, and restructure without directly touching the files. It's my working memory layer as a strategic advisor.

This week, an MCP server was added to it. And it's the first time I've had a real experience of sharing memory with a machine.

On Wednesday, I needed something mundane: a list of board members in a number of Danish companies. Not for a database. Not for a product. Just for an afternoon of lead research — who sits where, and which chairmen overlap with people I'm already talking to.

That kind of data is public. It's on datacvr.virk.dk. Every single company in Denmark has a page. It should take twenty minutes.

There's a difference between working with AI and working practically with AI.

The former is using ChatGPT from a browser. The latter is building something where AI is one step among many—where data needs to be fetched, validated, sent to a model, checked, transformed, and stored elsewhere. Almost all commercial AI work is of the latter kind.

And if you're doing the latter, Nushell should live in your terminal.

I recently wrote about how I automate my blog releases in one command — text in, article with image out. That flow runs in a shell most people have never heard of: Nushell. And it's not a random choice.

When I build what I call compound AI systems — meaning systems where eg. language models and image models are combined with other components to solve a real task — the most important question isn't "which model are we using?". It's "what do we surround the model with?".

And that's where Nushell comes in.

I talk a lot about AI solving real business problems, not just impressing in a demo. So let me show you an example from my own daily life.

Every time I write an article for docujai.com, it has to go through the same chain of small tasks: post the text to Drupal via JSON:API, generate an illustrative image, give the image the right visual expression with my logo in the corner, upload it to the article, set alt-text, patch the relationship between node and file. It doesn't take long to do manually. It takes a long time to do it every time.