I read all of Claude Code’s ~1600 line default system prompt top to bottom while studying @mattpocockuk course. Here are the most interesting things I found. 1/ If your codebase is slop, you’ll get more slop... 🧵👇

Humans and our ancestors have been more intelligent than Mosquitoes for 200 million years… … and we still haven’t managed to make them extinct. Think about that. Life is resilient.

I've seen a lot of Active Directory environments in the last 5+ years. These are some of the most dangerous issues I recommend reviewing and addressing in your environment. Treat it like a check list. If you have 0 of these, you're doing a great job.

Every time I see these posts I encourage people to follow everyone in the replies. There’s definitely people on here doing this, we have to build our circle of followings so we can share with each other! I’m doing two techniques for this: 1- Test homelab box (no private data) where I just give AI direct access (currently grok bot) so I can see all the mistakes and problems this can cause (like grok bot using my PAW for a Tailscale relay - see image…), plus how fast it can work without constraints. 2- At work, where I build systems around CNI, I’m using Infrastructure as Code (Terraform, Ansible) with security gates/tiering via pipelines (GitLab). Two extremely important components of this: - No AI agent has access to privileged credentials - these are gated behind GitLab protected branches and human review. - Indeterministic AI actions are converted into deterministic instructions by writing the IaC - it will do *exactly* what is in the runbook/Ansible role/Terraform plan. This solves the risks at a fundamental level whilst granting the productivity gains from AI. I can deploy and manage stuff super fast because the AI can do all the research on syntax and write the IaC in minutes. But I still have total control, and any changes to infrastructure are because I said “yes”. It’s actually beautiful because we get to ride the coattails of how good AI is at coding… Screenshot examples of both below (note, all on my test lab, I use both techniques here for experimentation)

The common meaning (not definition) of software is changing. Most people think of it as “an app” - a persistent tool that you can reuse. @sporadica is right, it’s unlikely everyone will build their own apps. However, nearly everyone using AI today is building their own software. You ask AI to analyse a spreadsheet, it writes a python script (software) specific to the task, runs it, and gives you the analysis. The python script is binned. In time this loop gets tighter and tighter (and faster), where software is constantly built on-the-fly to solve individual problems, by the AI, as requested by the human. The human doesn’t even need to be aware that somewhere in the middle, they built software. There also might be a future where AI generates the most likely *pixels* based on your input, forming an on-the-fly user interface, specific to your needs and based on your UI preferences. Everything else becomes middleware. I don’t know what happens to SaaS at this point.

But do it still say “load-bearing”

Grok @bot installed Proxmox remotely over KVM (via an RDP jump box with Tailscale). Literally interactively using the GUI/console of the server. Viewing the screen, clicking/typing. It’s one thing for AI to send CLI commands. Physically controlling a computer is amazing!

Woke up this morning to exactly what I asked for from grok @bot - a map of my test homelab with history. All I did was ask grok bot to login to my lab with Tailscale, and start looking around! Super cool