Franz Hemmer

AI Evangelist / Principal Software Engineer
Creator of HemSoft Developments
-- Ships code, not slides -- I'm new here - be nice :)

[Franz Hemmer @franzhemmer]

@theo Last I checked the Windows version just is unusable.

[Franz Hemmer @franzhemmer]

Opus 5 is just rubbing me the wrong way. My long term observation is that GPT 5.x models for me have worked really well because they follow instructions much more rigorously than Opus models.

The way I discovered this was when we got introduced to skills, I created a skill-creator skill that would write up all skills so that they at the end had a protocol skill they had to consult before concluding their work. That way I have the ability to guide every skill to do something uniform like logging its use etc.

This never really worked very well as I was using Opus models at the time. I'd say protocols were consulting half of the time. When GPT 5.4 Codex came around I gave it a go and noticed that it never missed a beat when it comes to the protocols. That got me into GPT being the main driver. I need that trust.

So when Kimi K3 all of a sudden is able to do something GPT 5.6 was struggling with, that impresses me. Hopefully it will do the protocols! :)

[Franz Hemmer @franzhemmer]

@burkeholland Yes!

[Franz Hemmer @franzhemmer]

@unclebobmartin @CaryKelly11 In Denmark you can get a PhD for free. Over here you have to be rich or put yourself in debt to do the same. Go figure … Now go forth and flame all you want on your political views. Facts stay the same.

[Franz Hemmer @franzhemmer]

@cassidoo Why are you not walking while doing that? You're amazing haha!

[Franz Hemmer @franzhemmer]

Try to reduce the wiggle room by encouraging the AI/LLM to be critical of your prompts and remove any ambiguity by explicitly calling it out in an AGENTS.md or your preferred agent or however you want to inject this small footprint into your every prompt instruction.

Kind of like encouraging a small 'grill-me' injection into the prompt.

[Franz Hemmer @franzhemmer]

@unclebobmartin Thank you for teaching me that mutation testing is a thing!

[Franz Hemmer @franzhemmer]

@unclebobmartin Time to open your mind to Zig Bob!

[Franz Hemmer @franzhemmer]

@CodeWithAmann VSCode for the win

[Franz Hemmer @franzhemmer]

@burkeholland My religion luckily only prevents me from going back to C++

[Franz Hemmer @franzhemmer]

If you pay attention to the wisdom of @unclebobmartin and implement attention to CRAP, mutation testing and general quality, security, and testing coverage principals both at the context level (AGENTS.md/skills/etc.) and the CI gating level, your codebase will have a much bigger chance of succeeding long term if that is your ambition.

[Franz Hemmer @franzhemmer]

I'm trying out the Codex Pro $100 plan, and two weeks in I can not max this thing out. I'm living the dream right now, running everything with GPT 5.5 xhigh (regular context window, non-fast) and I'm loving it. I wish I could use Copilot CLI instead of Codex CLI but there are pros/cons to both. I love the /goal feature btw.

The equivalent usage knowing my usage very well is estimated at ~4K/month on the GitHub Enterprise plan at work. The difference is nuts.

[Franz Hemmer @franzhemmer]

@burkeholland @github @mattpocockuk Won't be long before I can tell my grandkids stories about how I used to write code entirely by hand and they won't believe me.

[Franz Hemmer @franzhemmer]

Anyone else getting this on first session request?

Error: Execution failed: Error: 400 "checking third-party user token: bad request: Personal Access Tokens are not supported for this endpoint\n" (Request ID: E599:3D91FF:607E4C:690DC2:6A075036)

I had copilot scrutinize that there is no trace of a PAT anywhere and that I'm authenticating correctly with OAuth. No issues in my account setup - all looks green and connected.

[Franz Hemmer @franzhemmer]

@OrenMe I can blow through $200 of inference in an afternoon - not sustainable for personal use at all. I totally understand that the premium request model is not sustainable but it sure is a splash of cold water in the face to wake up to the token-based pricing coming up.

[Franz Hemmer @franzhemmer]

@GergelyOrosz @ahejlsberg Anders du er en dansk helt! ❤️

[Franz Hemmer @franzhemmer]

@burkeholland @github Oh lordy, its already in there! Woot!!

[Franz Hemmer @franzhemmer]

@burkeholland @github Looks beautiful - now if only there was way to get a Terminal experience in there with each project session that would be amazing.

[Franz Hemmer @franzhemmer]

@awakecoding I'm on Enterprise and I'm an org admin - I can't see this report anywhere. Any more details on how to get to it would be appreciated.