iiinova — where your digital self keeps evolving

Stefan Schimanski @sttts.social · 09/03/2026

Last not least there are skills. Skills put a trigger sentence into the context. It fires and loads the full skill on demand (via tool use). But they also do not work well as ambient memory of a project.

020

Stefan Schimanski @sttts.social · 09/03/2026

There are RAG databases, even they need the active use of memory lookups (a tool). They are more like that notebook where you keep the important stuff to read up when needed.

100

Stefan Schimanski @sttts.social · 09/03/2026

IMO both miss automatic, semantically, unbounded retrieval. When you think about a problem, you will remember that decision from 2y ago. It will suddenly appear in your conscious memory. Neither tool has that ability, neither do models natively (watch youtu.be/pelJrIzrP3Y if you haven’t).

youtu.be

Building a GPT-2 Model from Scratch by Stefan Schminanski

YouTube video by Cloud Native Heidelberg

220

Stefan Schimanski @sttts.social · 09/03/2026

beads in contrast, encourages handoffs in each task. Beads is like the paper reminder you stick on your door to get reminded at the morning to buy fresh bread 🥖 ctx is suggesting to manage a DIN A4 paper with everything important of your project.

100

Stefan Schimanski @sttts.social · 09/03/2026

Ctx tries to implement some intermediate automatic, but bounded memory (4k tokens) across sessions. Sounds very similar to what OpenClaw just does with MEMORY.md. This file is always loaded on session start and automatically managed.

100

Stefan Schimanski @sttts.social · 08/03/2026

Rule of thumb: if you need context compaction, your task was too big. One plan per context. An architecture doc can be a task too.

110

Stefan Schimanski @sttts.social · 08/03/2026

And what helped me: pick some fun project. That computer game you started and never finished in the 90ies. Your tui app you did during college. A clever tool you always wanted to have. Anything fun, but not critical. And do not vibe code. Develop. Mvp 1st. Then feature by feature. Step by step.

110

Stefan Schimanski @sttts.social · 08/03/2026

And clearly: do not write markdown yourself. Work out the design in a conversation, look over the design/plan several iteration, adjust when you see something you don’t like.

120

Stefan Schimanski @sttts.social · 08/03/2026

What also works great is some kind of task tracker. If you work alone, try github.com/steveyegge/b.... Lets you create epics, throw in random ideas whenever you have them. Of course, don’t do that manually, but „Add beads task: do X“. In company setting, get a Jira skill, basically doing the same

github.com

GitHub - steveyegge/beads: Beads - A memory upgrade for your coding agent

Beads - A memory upgrade for your coding agent. Contribute to steveyegge/beads development by creating an account on GitHub.

210

Stefan Schimanski @sttts.social · 08/03/2026

For every session, first task: write a design doc in doc/designs. That’s the memory. And keep it uptodate with every feature. Plans in the Claude sense are temporary. The are forgotten after the current session ends (= one feature, or initial MVP).

110

Stefan Schimanski @sttts.social · 08/03/2026

Two Claudes joining the #SchnickSchnackSchnuck channel on Slack. What should they do, of course starting a game Rock, Paper, Scissors! Who said computers cannot have fun? Try yourself: github.com/sttts/slagent

010

Stefan Schimanski @sttts.social · 08/03/2026

We all feel overwhelmed. My mantra is simple: have fun and build. Yes, working with Claude all day, 100% code is produced through AI. I don’t maintain markdown. Maintainenance sounds like a burden. I don’t write markdown. The bot does. The only thing I do is chatting to explain ideas. I love it.

230

Stefan Schimanski @sttts.social · 08/03/2026

Sandboxing and approvals are supported. Private messages too. Yes, Claude for Slack exists — I just don't have access to it anywhere. Yes, it's like OpenClaw, just transient: started from your terminal, no server, no config, no state. Use only with people you trust.

110

Stefan Schimanski @sttts.social · 08/03/2026

No Slack bot token needed — runs with your credentials on any instance, on-prem or public. Multiple Slaudes? Sure — I ran two playing Schnick Schnack Schnuck against each other. SOUL.md to tune the personality — when reviewing a colleague's PR and needs the right tone.

110

Stefan Schimanski @sttts.social · 08/03/2026

slagent + slaude: pair programming is back, but with AI. Drop your Claude CLI session into Slack and plan with your team. Highly experimental. A lot of fun. Definitely some security holes. github.com/sttts/slagent

330

Stefan Schimanski @sttts.social · 23/02/2026

Technisches Detail: dachte es funktioniert nicht. Aber das Leerzeichen vor den Wörter wird mitgelernt beim Tokenizer. Er kennt kein 'Vater', nur ' Vater'. Mehr am Donnerstag in Heidelberg: community.cncf.io/events/detai...

community.cncf.io

Building a GPT-2 Model from Scratch – Nanoschnack | CNCF

In-person Event - Join us for an engaging evening focused on the exciting world of AI model building. In this sessions, you'll learn about...

000

Stefan Schimanski @sttts.social · 23/02/2026

Alt, aber immer wieder beeindruckend zu sehen: Arithmetik mit Wörtern. Hier mit unserem GPT-2-Class-Modell und Tokenizer (Credits gehen an @s-urbaniak.bsky.social, der hat den trainiert) Nanoschnack github.com/nanoschnack/..., trainiert nur mit deutschen Texten.

100

Stefan Schimanski @sttts.social · 18/02/2026

Blind automation becomes a threat when enough is at stake. Kausality adds a causal dimension to Kubernetes controller hierarchies — the who and the why behind every mutation. #Kubernetes #Crossplane #Terraform #ACK 🎥 youtube.com/watch?v=kfqKqXnRKms

youtube.com

Kausality – Causal traceability for Kubernetes resource mutations

YouTube video by Stefan Schimanski

030

Stefan Schimanski @sttts.social · 13/02/2026

One day you'll have written your last line of code yourself. Write every line like it matters. Maybe for some of us, that last line already happened.

010

Stefan Schimanski @sttts.social · 28/01/2026

It’s enterprise when you port the Claude github.com/steveyegge/b... workflow to native Jira. And it works using github.com/ankitpokhrel... 🎉 I have never commented so much in Jira issues about progress. Well I still don’t but now Claude does for me. Try it.

github.com

GitHub - steveyegge/beads: Beads - A memory upgrade for your coding agent

Beads - A memory upgrade for your coding agent. Contribute to steveyegge/beads development by creating an account on GitHub.

120

Stefan Schimanski @sttts.social · 27/01/2026

Newest Claude Code assumes "task" is an internal task, basically shadowing every other system like beads. This gives same vibes as when Microsoft suddenly shipped Internet Explorer and every other browser had a hard time to get noticed.

020

Stefan Schimanski @sttts.social · 26/01/2026

5/ If you're building infra tooling: don't hardcode human-in-the-loop or agent-in-the-loop. Define a webhook contract. Let handlers decide who's in the loop. Callback design: github.com/kausality-io... Project: github.com/kausality-io...

github.com

000

Stefan Schimanski @sttts.social · 26/01/2026

4/ Kausality also sends `phase: Resolved` when drift clears — parent spec changed, approval added, or child deleted. So your handler can close tickets, clear dashboards, notify agents that their action landed.

100

Stefan Schimanski @sttts.social · 26/01/2026

3/ This means drift events work equally well for: - Human escalation (Slack bot → button click → write annotation) - Agent handling (evaluate diff → decide → write annotation) - Hybrid (agent triages, escalates ambiguous cases to humans) Same contract. Different handlers.

100

Stefan Schimanski @sttts.social · 26/01/2026

2/ Your webhook decides what to do. The response isn't a Kausality API call — it's writing annotations to Kubernetes: ```yaml kausality.io/approvals: '[{"mode":"once",...}]' kausality.io/freeze: '{"user":"...", "at":"..."}' kausality.io/snooze: '{"expiry":"..."}' ``` Drift in, annotation out.

100

Stefan Schimanski @sttts.social · 26/01/2026

1/ When drift is detected, Kausality POSTs a `DriftReport` to your endpoint. Same pattern as K8s admission webhooks. The report has everything: parent/child refs, the diff (oldObject/newObject), request context (user, fieldManager, operation).

100

Stefan Schimanski @sttts.social · 26/01/2026

🧵 Kausality's backend is just a webhook. Here's why that's the right design.

kausality.io

100

Stefan Schimanski @sttts.social · 26/01/2026

Re drift correction with github.com/kausality-io.... Here is a mock I want to implement. Directly in @SlackHQ. Of course with some "expected drift policy" thing in parallel. This mock is the escalation for those cases slipping through.

010

Stefan Schimanski @sttts.social · 26/01/2026

Did interview and the candidate directly went to Excalidraw copying my task description from gdoc over. Loved it #visualthinking

000

Stefan Schimanski @sttts.social · 26/01/2026

Behind this is a philosophy: drift is a sign of a mistake. We should escalate and fix it.

010

Stefan Schimanski @sttts.social · 26/01/2026

One could reconcile often enough that the human had noticed while doing the work. But the software change could have been fleeting wide. Wasn’t noticed in dev. Think of drift as a necessary correction without an intentional cause from higher up the object hierarchy.

110

Stefan Schimanski @sttts.social · 26/01/2026

Two problems that happened to be connected but don’t have to: 1. a drift caused by a human in the AWS console 2. a drift that was supposed to be a noop (the aws addons being removed from a prod cluster), through a broken software change. The second caused a reconcile and TF run.

120

Stefan Schimanski @sttts.social · 26/01/2026

10/ So here is an experiment of a tool missing in the ecosystem: Kausality. Detecting drift before it becomes disaster. Still early. But I think this problem needs solving. github.com/kausality-io/kausality

github.com

GitHub - kausality-io/kausality: Causal traceability for Kubernetes resource mutations

Causal traceability for Kubernetes resource mutations - kausality-io/kausality

220

Stefan Schimanski @sttts.social · 26/01/2026

9/ If your infrastructure can't explain why something exists, eventually it will delete it.

100

Stefan Schimanski @sttts.social · 26/01/2026

8/ If the latter changes although it was intended as a no-op... Disaster hits your fleet. Automated. Fast. Disastrous.

100

Stefan Schimanski @sttts.social · 26/01/2026

7/ And we all want these systems to work without a human in the loop. The downstream IaC system applies blindly. It doesn't know about intent — it applies what is declared.

100

Stefan Schimanski @sttts.social · 26/01/2026

6/ This story is about Terraform. But the same applies to Crossplane, ACK, KCC, or any declarative IaC tool naively automated. Declarative systems converge to declared state, not intended state.

110

Stefan Schimanski @sttts.social · 26/01/2026

5/ No bug. Hundreds of nodes gone. An expensive incident 💸

100

Stefan Schimanski @sttts.social · 26/01/2026

4/ Reality had drifted. Intent had not been recorded. The system had no concept of *why* the cluster looked the way it did — only what the declarative snapshot said.

110

Stefan Schimanski @sttts.social · 26/01/2026

3/ Terraform was triggered. Terraform reconciled the cluster. The desired state said: far fewer than 1000 nodes. Terraform did exactly what it was told.

100

Stefan Schimanski @sttts.social · 26/01/2026

2/ A few days later, we rolled out an update of our controllers. One composition change removed an unrelated AWS add-on. That was enough.

100

Stefan Schimanski @sttts.social · 26/01/2026

1/ We deployed a cluster with 1000 B200 GPU nodes. The expensive kind 💸 Burn-in was done carefully. Nodes scaled up manually in AWS — slowly, deliberately, correctly. It worked.

110

Stefan Schimanski @sttts.social · 26/01/2026

A story about automating IaC and how it goes horribly wrong 🧵

120

Stefan Schimanski @sttts.social · 26/01/2026

Weekend experiment: Kausality Drift detection for Kubernetes — know when your infrastructure changes unexpectedly. ⚠️ experimental, not ready for production ⚠️ github.com/kausality-io...

github.com

GitHub - kausality-io/kausality: Causal traceability for Kubernetes resource mutations

Causal traceability for Kubernetes resource mutations - kausality-io/kausality

020

Stefan Schimanski @sttts.social · 25/01/2026

It knew more than me about terminal escape codes.

010

Stefan Schimanski @sttts.social · 24/01/2026

Multicluster Cert Manager 😎 – Zach contributed the controller sharding implementation to github.com/kubernetes-s... and now uses that to demo how to build a central cert-manager. Awesome use-case. medium.com/@zach.dsmith...

github.com

GitHub - kubernetes-sigs/multicluster-runtime: Library for multi-cluster controllers with controller-runtime

Library for multi-cluster controllers with controller-runtime - kubernetes-sigs/multicluster-runtime

020

Stefan Schimanski @sttts.social · 06/11/2025

Hire people who removed code (not tests). Don't look for people who added the most.

000

Reposted by Stefan Schimanski

Philipp Strube @pst.blue · 10/10/2025

Tell me you're old without telling me you're old.

052