Becoming Operational - Why I Switched from Mistral to Claude
Part 1 of 4 in Becoming Operational
- Becoming Operational - Why I Switched from Mistral to Claude
- Becoming Operational - The Conventions
- Becoming Operational - Our Own Vocabularies, Linked to the AAT
- Becoming Operational - Arslantepe Is Online
Hi again! It’s been quiet on the blog for a while—but not in the repositories. A lot happened, and over the next few posts I’ll tell you how LOD-AM went from a hand-written example to its first real dataset online. But before I get to that, there’s something I want to get out of the way. I promised to be transparent about how this project is built, so here it is: I changed my AI assistant.
From Mistral to Claude
In How I started, I wrote that this website was pretty much a vibe coding attempt with Mistral. That was true back then. Since then, LOD-AM has been built with Claude, Anthropic’s AI assistant, working directly in my GitHub repositories.
I am glad enough to work with a lot of colleagues, who love to experiment with every piece of technology available. And when I got a recommendation and a report of his recent progress from my dear colleague Lasse, I just had to try it out. We'll see how the European competitors of Anthropic will do. I am hoping for the best. Competition boosts business afterall.
Why? Time.
Here’s the thing: LOD-AM is a free-time project. It has to fit into spare evenings between everything else. Every hour I spend on it has to count—so I have to be as effective as possible. And with Claude, I simply get more done.
What made the difference for me isn’t that Claude can write code. It’s that it can carry a task from start to finish: read my conventions, change the data model, run the checks, look at the result in a browser, and come back with a pull request I can review. And when something fails, it tells me what failed—instead of pretending it worked.
Who decides what?
This part matters most to me, so let me be clear about it:
The research decisions are mine. What a value in a publication means, which type an object belongs to, which period a context dates to—whenever something is a domain judgement and not a fact from a specification, Claude doesn’t decide it. It writes the question down, and I answer it. My sandbox, still my rules.
Claude does the heavy lifting. Writing and testing the conversion, looking up ontology terms, checking every AAT identifier against the Getty endpoint, building the pages you see here, keeping the documentation up to date.
Nothing goes live without me. Claude works on branches and opens pull requests. I review them, and I merge them.
Trust, but check
There’s one mistake an AI assistant makes all too easily in a project like this: inventing an ontology term that looks perfectly plausible but doesn’t exist. That’s why every single term in LOD-AM is now checked automatically against the published ontologies before anything is merged. I don’t have to trust the assistant—I can check it.
And these posts?
Full transparency here as well: Claude drafted this series from our notes, issues and pull requests. I went through it, changed what didn’t sound like me, and I stand behind every word.
Curious how the rules for Claude look? They’re in the repositories, right next to the code:

View LOD-AM/knowledge-graph on GitHub →
As always, if you have thoughts, suggestions, or just want to talk about working with AI in research, you know where to find me: info@lod-am.net or (if you have a GitHub account) in a comment below.
Comments
Comments are provided by giscus and stored as GitHub Discussions. Loading them connects your browser to giscus.app and GitHub, Inc. in the USA and transfers your IP address and browser data there. The USA does not have a level of data protection equivalent to the EU's; US authorities may access the data. By clicking the button, you consent to this transfer (Art. 49(1)(a) GDPR). Details: privacy policy, section 5.