Engineer Manager at TELIS/GWVS. AI, Cloud-Native and .NET, Speaker, Hobby Shouter and Guitarist @not_brawn
Regensburg, Germany
Joined November 2009
- Tweets3.7K
- Following468
- Followers337
- Likes11.9K
Forget Vibe Coding, do Guided Coding!
1) Plan iteratively with the Coding Agent
2) Let the Coding Agent implement the plan
3) Review and iteratively guide the Coding Agent towards better design, security, performance, etc.
#AI #CodingAgents #GuidedCoding #VibeCoding
I really liked Opus 4.8 and Opus 5 for their rigorous criticism when reviewing things. When writing docs or plans, they produced texts that were too verbose and too pompous. Which seemed elaborate to me at the beginning, became annoying after a few time. Glad that's gone.
Cool! Would you happen to know how much of your weekly usage was used up by this? I guess you are on the 200$ / month plan?
I'm on the 20$ plan - and Astra flies through my 5h limits within 10 - 15 minutes.
Kenny Pflug retweeted
Dynamic web parsing in .NET just leveled up. ⚡ AngleSharp.Wasm v1 extends AngleSharp with WebAssembly execution capabilities—joining AngleSharp.Js to handle modern, client-side dynamic web apps without spinning up heavy headless browsers. Lightweight and fully spec-compliant. 🚀
If you happen to speak German, check out the newest video on @BFreakout's YouTube channel where I was invited to talk all about #GuidedCoding: youtu.be/-jQvlwiCdLw?is=9zRv…
#AI #CodingAgents
Kenny Pflug retweeted
I just proposed an official @leanprover 4 formal semantics for C# - giving C# a foundation for machine-checkable proofs.
AI agents are getting very good at writing code - and very good at breaking it.
In the agentic era, "tests passed" won't be enough for critical software. We need to be able to prove properties of the code we ship.
And this is becoming practical.
Rust already has Aeneas -> Lean. Microsoft Research is using it to formally verify SymCrypt. Java and C have roots into formal verification too.
C# should not be left behind.
AI can write the code.
AI can construct the proof.
A deterministic proof kernel checks whether the proof is actually correct.
That is a very different trust model from:
AI writes code -> AI reviews code -> hopefully both were right.
Proposal:
github.com/dotnet/csharplang…
Nicht vergessen: Heute Abend My Coding Zone Twich Stream! Als Gast bei @BFreakout stelle ich #GuidedCoding vor. Let's go!
#AI #CodingAgents
meetup.com/de-de/my-coding-z…
Interested in my #GuidedCoding approach for Coding with AI Agents and happen to speak German? Join @BFreakout and me on the upcoming Friday at 8:30 PM CET: meetup.com/de-DE/my-coding-z…
#AI #CodingAgents #csharp #dotnet
🦾 Of course, this was largely implemented by #AI #CodingAgents, mainly Claude Code with Opus 5 High Reasoning and Codex with 5.6 Sol xhigh Reasoning, following my #GuidedCoding approach.
kenny-codes.net/docs/guided-…
🚀 Light.PortableResults 0.7.0 is out now!
📈 Overhaul of the Metadata System, serialization got a 16% to 55% performance boost, new UUIDv7 validation check.
#csharp #dotnet
github.com/feO2x/Light.Porta…
I don't know what @AnthropicAI did in the past week, but Opus 5 is now way less petulant and touchy when its initial goals are resteered. Way more usable in high-interaction conversations, like when planning or reviewing.
#AI #CodingAgents
This is what happens when my wife is away for the weekend with her girls...
#AI #CodingAgents
ALT My Claude Code Usage today. Almost 175 million tokens. I would have to pay rougly 900$ if I wouldn't have a Claude Pro Subscription.
Interested in the results? Checkout github.com/feO2x/Light.Porta…
Give your #AI #CodingAgent feedback loops. Not only compilers, linters, and automated test suites, but also automated benchmarks. They even write cool reports for you!
#csharp #dotnet
ALT A screenshot from claude code, reporting the results of BenchmarkDotNet runs. Overall, performance improved between 16% and 55% in difference scenarios.
Also, having multiple teams on the same map is great for finding new online friends. Great game design!
@GearsofWar E-Day looks fantastic! Horde Siege seems like a great overhaul - the old 50 wave format was cool, but also grindy. The mission-based approach now looks like it's working great. Can't wait for the beta next week!
youtu.be/2-710kEZvK4?si=dad0…
But when steering is so hard with Opus 5, this simply takes up your time. I still use it for reviewing plans and implementations, but for actually planning and implementing, GPT-5.6 Sol is my go-to model.
In the end, it took me more time to write plans. I still like the aspect that Opus 4.8, Fable 5 and now Opus 5 are super-critical - I need counterparts that challenge my ideas. GPT-5.5 and GPT-5.6 show less of this behavior.
I wrote a plan following my #GuidedCoding approach, and I wanted things to be done differently than Opus 5 anticipated. This resulted in a degraded plan: dulled words, incomplete descriptions, as if the model had lost interest.