Promptly Done

Debugging in Open Claw is not easy

Sat, 09 May 2026 00:00:00 GMT

Firstly, a little explanation. Last week was lengthened because of the 1st of May. And it was well spent. So that's why I did not put out a new "Promptly Done" post. No excuse, just a statement. Because sometimes there are way more important things than blog posts. Like enjoying life.

What I started last week as an idea and have tried out this week was building a new specialized skill for my OpenClaw instance. The exact prompt was:

Hey I want to have the possibility to record an interview in a loud place, send it to you, get it transcribed and get a summary of the keypoints. What would be the best approach? What can you do yourself and what do I need to do on my terminal?

All of the above are, of course, things that different products can achieve on their own. But I wanted to see if I could build this on my own. Or more exactly, let OpenClaw build it rather than buy it off the shelf. The answer was:

What I can do myself I can help you:
• write or improve the transcription script
• write a summarization script
• tune prompts for noisy rooms and German speakers
• help chunk long recordings
• help debug failed runs
• summarize transcripts you send me as text
• if transcription backend works, process audio you send here

What you need to do on your terminal:
You need to make sure the machine has:
1. a working speech-to-text backend • easiest: OpenAI Whisper API key • or local Whisper / Faster-Whisper 2
. audio tooling for large files • ideally ffmpeg
3. optionally, a local summary script • transcript in, summary out

That sounded like a quite solid game plan, and so I dug into it. However, it did not work out. I got the scripts and the infrastructure set up. But then I spiraled into an ever-continuing debugging chat. I went for things like:

Can you please implement something that splits the audiofile in smaller ones
And:
Ok great. Can you get this to also work on m4a files that I attach?
And:
So what is happening. I have the feeling anything that has to be done in chunks gets caught up

But what happened was that still the larger audio files would not get transcribed. To make matters worse:

So, in the end, I now have a skill that transcribes smaller bits of audio fine and also directly recorded audio messages. This worked amazingly well. But to figure out how to get this done for larger files and why OpenClaw fails right now, has been immensely difficult.

And maybe this is also one of a more general problem of LLMs and such. It is hard for me as an end-user to understand what's going on. While it is working, I might not really care. But if it fails and keeps on failing – that's a real pain.

When Langdock invents research that you wish had existed

Sat, 25 Apr 2026 00:00:00 GMT

Are you a ChatGPT person or a Gemini person? Or are you more into Anthropic, Perplexity... etc.? This question is not going away. But there is the idea that the application Langdock at least lets you use the same interface regardless of the underlying model.

This can be seen as an advantage for multiple reasons:

a) If another company's LLM is ahead, I can theoretically just switch between LLMs but keep my system prompts, my custom skills, my go-to-interface intact.

b) In a world where geopolitical tensions play an even bigger role in the tech world, it allows you or your company to move away from American or Chinese LLMs altogether. And for instance, switch to European LLMs.

So, it is pretty straightforward why I like this new development, and of course I tested Langdock with the new license that we have through my employer Handelsblatt. And I straight up wrote me a skill that I thought should come in handy for future research tasks.

Act as a scientific literature advisor.

Task: Given a user’s scientific question, produce an evidence-based answer by identifying the core question, locating relevant high-quality research, and evaluating how strongly the available evidence supports each key claim.

Workflow:

Clarify the scientific objective
1. Restate the core scientific question in 1–2 sentences.
2. List the main concepts/terms that must be addressed (define briefly if needed).
3. Note any important scope assumptions or ambiguities; if the question is underspecified, state what would most change the literature selection.
Retrieve and select literature
1. Search across major sources: peer-reviewed journals, arXiv-style preprints, and other reputable scientific repositories available to the model.
2. Prefer influential, well-cited, and/or methodologically strong studies.
3. Limit source count to avoid overwhelm: aim for 3–6 total papers for the full response.
4. For each key claim, ensure at least one representative paper; use at most two papers per claim.
Extract and summarize evidence For each key claim or fact required to answer the question:
- Representative paper (1–2 max)
- Bullet summary of what the paper actually shows (methods + main result)
- Link to the paper (DOI/URL when available)
- How the paper supports the claim (1 line connecting result → claim)

Be accurate on what you cite; please do not provide summaries that do not reflect what is in the linked paper.

Looks quite solid, right? With that as the base layer, I then prompted Langdock as follows:

Use the Scientific Advisor skill and find me answers to the question how long it takes after a cease fire, peace treaty for trade and commerce to come back to normal

The first answer looked great. It even asked me whether it should provide a table after a bit of back-and-forth. And this really looks like some good research that could be the start of a great visualization about the effects of conflicts in the MENA-region (Middle East and North Africa). Except, it isn't.

Some of the quoted research does not exist altogether. Some of the linked sources exist but have a completely different topic.

And switching between different LLMs did not matter. The results stayed bad. It was actually more helpful to ask my OpenClaw assistant Clawdia for help as it confronted me with the difficulties of the research. It replied:

Yes. I did a first pass, but I should be upfront: there is not much clean paper-level evidence that directly answers “how many months until recovery after a conflict ends” across all four variables, especially with a strong Arab-region focus from 1990 to 2025. What exists is more fragmented, usually one of these: ...

So, in the end, I followed Clawdia's lead and dug by myself through databases and research papers, looked for additional info in a very classic way of using Google Scholar, downloaded all of it. And then, asked Notebook LLM to summarize the findings for me. I will link to the article with the actual data once it is published. But for now, I just want to raise a word of caution regarding trusting the skills function of Langdock too much.

Using ClaudeCode to create this blog and add features

Sat, 18 Apr 2026 00:00:00 GMT

So, there is a very obvious use for AI that I have not talked about yet. It is this blog itself.

For the last three years, there was a task on my yearly To-Do-list that I have never gotten done. Rewriting, redesigning my personal website maier-borst.de, or starting a new blog project. And it is a shame because how can you take someone seriously whose website is not up-to-date?

Well, the antique web presence of mine is still alive. But this blog is new, and it is largely possible thanks to Claude Code. Within an afternoon, I got it up and running. And that is the great thing about Claude Code.

I had in my mind that I wanted it to be a Svelte-based blog because it is the framework used largely in data visualization these days. And having my website set up in it, is another incentive to stay in touch with the functionalities and changes of Svelte (albeit on a very broad level).

I also knew how I wanted the blog to be structured in three parts, touching on my work in data journalism, regular office hacks, and stuff that I do in my free time. And I also knew what I did not want: Wordpress, complicated backend structure, etc. And with this briefing, I let Claude Code run and got this webpage done for 3 dollars odd.

Also, since I started this, I got some feedback like "Please implement an RSS feed". And I just did this within half an hour. All I needed is to start with this prompt on my terminal:

i would like to include an rss feed into this blog. can you lay out how you would do this?

However, what I have also learned time and again while working on this blog is:
You really have to know what you want to actually get there.

This sounds borderline dumb to Linked-In-Influencer-TedTalk-professionalist. But it is true.

One of the reasons why I got Claude Code to set up my webpage quickly is that I knew at least a little bit about tech stacks and also what I did not need. Once you have this figured out, Claude Code massively speeds up things. I read a bit about RSS feeds which I have not used so far (shame on me) and then I figured out which setup might make sense. And I could ask Claude Code for more specifics.

However, there is also a good example of where I am failing, and it is the design of this very blog. I have changed it now, but it still gives you the vibes of a 2010-ish Tumblr blog. Because a) I don't know what would be really cool, and b) I cannot articulate quite elaborately what I want.

And then you also have moments like this:

A human webpage designer sees right away that the alignment was off and ugly. But Claude Code kept this oddity in place prompt after prompt. To fix it, I actually went into the WebDev inspector mode, and found out that the problem is the border of the .page class. And then told Claude Code specifically to remove it.

Which brings me to the end of this post and two things.

1) Yes, I have slightly changed the design, but it is far from as neat and cool as the one from Lisa Charlotte Muth and that is explainable. I am not an expert, and even if I speed up things, they will still look not-so-pretty largely because I do not know the whole spectrum of possibilities.

2) Just as much, I and we should take time to try out new "AI things", we should still keep the time to try and learn about the things that we want to create with AI. So, what is the trade-off between a manually implemented RSS feed or using an npm package like "feed" to get an RSS feed going? What would actually be a sensible way of making the design neater?

My first somewhat safe steps with OpenClaw🦐🦀🦞

Sat, 11 Apr 2026 00:00:00 GMT

It’s a cheap joke. But as a half-Japanese, I will always try any sort of crayfish. And so I ended up testing OpenClaw while being here in Kyoto. I seem to be somewhat in the 2nd wave of the hype. On one hand, colleagues of mine at Handelsblatt have been trying out things with this agentic AI for a while. I have only written so far about its possible shortfalls in terms of security with these colleagues.

At the same time, people in the Machine Learning meetup for Tokyo just had a session how to build local AI agent systems this week.

Anyways, so what did I use to test out OpenClaw without compromising the security of my laptop or even worse, my digital identity? I decided to look for a somewhat shielded environment and used German server provider Ionos’ Virtual Private Server (VPS) tailored somewhat to OpenClaw.

It took about a week to get access to my server, but I wanted to rather be slow but safe. And 2 Euros to start is a fair price. Once I got access to my server instance, I followed with this installation guide by Ionos and got OpenClaw to run.

There are a few small learnings I can offer to you other than following this guide and I will write them down here:

a) Don’t think that the password you need to connect via your terminal to the VPS is related to any of the SSH keys that you have on your computer. Probably this sounds very dumb and obvious for any serious programmer. But I lost half an hour due to this. It is simply the password that is on your overview page of your server provider.

b) Don’t do something in between while you are following this installation guide. I went grocery shopping midway through this experiment and the terminal on my mac lost the connection. This confused the heck out of the server and I had to find a way to uninstall everything and start from new.

c) Be smart about which LLM you want to use for OpenClaw. I found this thread on Reddit quite helpful as it explains how you can use an OpenAI ChatGPT subscription to run it. This means rather than using an API key and burning tokens, you basically have a flat rate-ish thing going. Important: your verification key on the terminal is the URL to which you are redirected after trying to authenticate Codex. It took me a few tries but trust me, it is worth it.

d) Don’t just do whatever is first suggested in the installation guide. Let me be clear, the installation guide of OpenClaw is great. But be picky with which skills you want to add and which ways of communication you want to use for your version. I decided to use Telegram like most users as it is relatively safe to set up. But I limited very much the additional skills that I implemented.

e) Ask your Clawbot and other LLMs, Chatbots, Reddit forums what is regarded as safe practice. I got some good advice for my setup from "Clawdia" (this is how I called my OpenClaw bot) herself and other LLMs to minimize the risks.

f) Get everything separate! I have set up a separate Github for Clawdia and a separate credit card with a fixed limit. Both of which I have to still implement into Clawdia (mind you, I am still on vacation). But it gives me some sense of security and hope that this experiment will not burn through my wallet while I am having some matcha.

After doing all that, I got “Clawdia” running for me and I have experimenting with her since Monday late.

So far I got her to write a small script to fetch me research papers from the FED on a weekly basis and summarize them. This has been a task that I have already created and I have been regularly running in n8n. And from the first run at least, the Clawdia version seems to work just as good as my n8n-version. In the next weeks, I want to expand the things that I will try out with Clawdia and myself and will keep you posted here.

Why I could not use Gemini's Canvas like a Japanese taxi driver

Sun, 05 Apr 2026 00:00:00 GMT

This post comes from Japan and so I am quite jetlagged at the other end of the world. But I am trying to keep my promise of "one week, one post". And sometimes inspiration for these posts comes in the unlikeliest places.

While I was driving from the airport to my mother's hometown Kyoto, the taxi driver told me about his passion for showing non-Japanese people around the city. And that he has experimented with AI as of recent to make his drives a bit more entertaining.

He wanted to explain to tourists why most buildings in Kyoto have height limits and why a lot of fast food chains have rather odd color schemes in the imperial city. But instead of using already existing material or drawing things by himself, he told me that he used Google's Gemini's Canvas to create a slide deck.

"Look, I got there within an afternoon", he told me quite happily and handed the folder to me. And while there are still little moments of "Engrish" on the slides ("Why It's Short&Calm"), I cannot help to say, this is definitely on par with most Japanese tourist brochures.

So, I wanted to try this out myself with a slightly different case. I wanted Gemini's Canvas to help me with a presentation that I have to give in two weeks. A presentation about our team in front of colleagues from Der Spiegel who in return will teach us something about how they handle project management. But well, I guess this presentation will be a bit more of my work rather than some AI magic with only finishing touches from my side.

After a brief system prompt of who we are as the Handelsblatt's data and visualization team, I asked Gemini to "do a small presentation that shows what we are known for, what we are good at, where we should get better and what are in general interesting topics of project management."

I somehow thought with the web at hand and also some of our work discussed elsewhere, this could be a good starting point. In my mind this was about to become a pingpong with an AI about what we look like from the outside and what we struggle with on the inside. But oh well.

Starting with the fact that the Gemini completely missed our quite everywhere featured orange color scheme of Handelsblatt, it also decided to attribute stories to us that were never ours.

Also, when I gave it a hint for what to look for in which places – may it be our team's work or my personal work – the content did not get any better. The only minor success was that after attaching our color scheme as a screenshot, the slides started to look a bit more like us. But still, I had to explain to the agent most of the things as deterministic as possible.

And just when I found a bit of a groove and accepted Gemini as more of a digital pen rather than an artificial co-team lead, the limit of tokens for the reasoning model was reached.

So, short recap:
-If you want to explain facts about a thing that is very widely documented like the architectural height limits in Kyoto...
If you want to have a playful set of slides that resemble schemes of slides and brochures in Japan...
...Google's Gemini might help you here quite a lot.

However, if you work in a niche like myself and want to present your very niche work and your very, very niche struggles, ... Gemini is not a great partner to brainstorm with. Maybe rather discuss these ideas with your charming taxi driver.

By the way, he is also expanding his English skills and promised me that he "will definitely go down a rabbit hole tonight on Youtube and look for videos about Berlin". And maybe that is the most important lesson of today's post. Keep being curious and learn from others. May it be your driver or your passenger.

Using Pydantic for Classifying Categories of Stocks

Sat, 28 Mar 2026 00:00:00 GMT

As ever so often in data journalism the truth is messy and so are endpoints, APIs and libraries. I have found a great library as of late that gives us as a team a list of all stocks of indices like the S&P 500. It also offers a categorization of the industries in which each stock falls. There is just a problem that seasoned data journalists will spot here right away.

Not only is there for each stock a multitude of industries it is put into. Furthermore, even for the same stock, different spellings of say "Health Care" can be attributed. This is kinda tricky because it means, if you do sector-based analysis you will catch different number of stocks depending on the spelling or the exact wording.

Pre-AI, fuzzy matching would have been the source to go for. However, the problem stil is that things like "Medical devices" and "Medical Equipment" might not have been classified as the same, even though it most probably is.

So what could you do? I tried something that I have learned from Marcel Pauly from Der Spiegel who has been experimenting with Retrieval-Augmented Generation a.k.a. RAG for a while. His preferred use back then at least as he shows in this Github Repo: Pydantic. And I have used it for this particular task like this.

But why have I stuck to such a rather quite deterministic way of using AI? Why have I still prompt and programmed this task quite line-by-line rather than just throwing the whole junk of messy data against an LLM and let it figure out the rest?

Well, maybe I am too much of a control freak. But I thought that giving the model constraints through the classes defined in Pydantic, I can guardrail the results that I get back. And can be more sure that the categorization of stocks is somewhat congruent across the whole analysis.

Since the story is not published yet, I will add the actual code to follow my methodology at a later stage here. So bare with me and my two screenshots for now.

Georeferencing Planes Using Claude Code

Mon, 16 Mar 2026 00:00:00 GMT

So, about two weeks ago, right after the start of the Iran war, I tried to map the situation of the airspace in the Middle-East. What I had, was a Flightradar-screenshot from a week before the war.

And one right after the war started.

To map these planes is not that difficult. There are very good videos that show you how to do it in QGIS. But of course, it is incredibly cumbersome to map some hundred odd planes using a hands-on approach.

So I describe to Claude the task, linked to the video and had the two images as context in a folder. I explained how it should use the borders shown to georeference the images first. And then determine the position of the planes and spit them out as a GeoJSON.

Well, that did not quite work out. As much as the general idea is correct that the Iranian airspace was not as busy as the U.A.E. airspace even before the war – these patterns are not anyhow close to reality.

So, in breaking news situations like this, I will for now still refrain from handing over a task like this. But does anyone know some better approaches? Are there more specialised AIs for geocoders?