Software engineer running Porsche service and restoration shop. And buildings apps for it.
Zilina, Slovakia
Joined May 2008
- Tweets66
- Following304
- Followers995
- Likes347
At this point it should be obvious, but I still get surprised from simplest LLM use cases. Claude with Basecamp CLI creating Porsche 911 engine rebuild report for the client. WTF.
Martin Adamek retweeted
Ok, how fucking cool is this.
stuartgarage.com/about
This small Porsche service and restoration shop in Žilina, Slovakia uses Basecamp to run the shop. And they even highlight it on their site!
This is why I absolutely love building software for small businesses. It's such a thrill to bring outstanding, affordable, best-of-the-best software to businesses like these. It's a joy to work for them.
Basecamp and Porsches. I mean... What a match!
cc: @dhh
Did anyone manage to tame @openclaw in multi-user environment? Agent per user. Separate user and team resources. Multiple teams on single machine. Testing such setup right now, about to deploy for the company as well as family.
Gemini 2.5 Flash surprisingly good at workshop manual parsing. Mixed orientations, illustrations, photos and yet the Markdown version is perfect. Good source for QMD and company’s OpenClaw.
I really to iterate designs with Claude's /frontend-design. I ask it to shoot 3 static designs at /ideas page and go from there.
Working on several apps in parallel created a new need for me - separate workspace for each app. I need a named display with terminal + browser. Would be much easier on Linux, but on Mac it is the AeroSpace by @nikitabobko that works for me.
github.com/nikitabobko/AeroS…
Having Rails dev environment locally is so annoying now that I tasted Tailscale + Tmux + Termius combo.
Nice, I finally configured this inline diff in LazyGit (and LazyVim). Much better. More important now with agent coding. Also standing desks.
Impressed by Haiku 4.5 - invoice parsing times down to half comparing to 3.5 and parsed all the test files correctly. Currently I can't find any invoice, that would not be 100% perfectly extracted. Also, simplified the prompt by [+15 -73].
Today's progress on tuning of LLM-based PDF invoice parsing. Now I can add models and sample files at any time and feel a little bit under control. Thanks ruby_llm for making model switching efortless.
Passing blob.path vs. temp_file.path is all the difference in perfect LLM invoice parsing. Using gemini-2.5-flash via ruby_llm. Is ActiveRecord messing the pdf somehow?
It is all about code reviews now. And I had to give a lot of feedback, not that Claude Code gets everyting perfectly on first try. Quite opposite. But it is so easy to navigate my virtual colleague now.
There has been such incredible progress in how Claude Code works with my Rails app since it was introduced, that I just can't believe it. In last weeks I did not write a single line of code. Not that I didn't want. I just didn't need.
For the LLM-based feature in my Rails app I added golden files with expected responses to my minitests. Now Claude Code can try different models and their versions using ruby_llm to create a perfect prompt. It is killing it so far for LLM invoice parsing.
Failed to setup @wafrisorg for my new app. Went with Rack::Attack. All new to me, but it’s crazy to see that shady stuff immediately in the logs. I was supposed to feel better, I guess.
Ouch, got some ai-coded feature commited automatically, probably by codex. I was prototyping the approach to some turbo streams refreshes and I switched to full auto. I didn’t realize it would also use git 🤦♂️. Caught in production crashing, reported by AppSignal 🙏.
About to release the Rails app we’ve been using internally for a while. What are the things that will bite me, security or resource-wise, that I can easily prevent?