Rudbek Design System v2 · Team update

AI Center of Excellence

Tools Update. The working tool set behind the CoE launch — what we use every day, why we chose it, and what the team should learn first.

AI Center key tools — hit the ground running

  1. 01
    GhosttyThe terminal

    Our daily terminal on the Spark and the Mac — fast, native, and built for long agent sessions.

    ghostty.org ↗
  2. 02
    HerdrThe agent control room

    Runs inside Ghostty and keeps every agent visible — one terminal for the whole herd.

    herdr.dev ↗
  3. 03
    Matt Pocock SkillsThe skill library

    Reusable agent skills for real engineering — the collection many of our installed skills come from.

    github.com/mattpocock/skills ↗
  4. 04
    GPT-5.6 Sol + Opus 4.8The frontier pair

    The hosted heavy lifters we relied on before Kimi — with Fable in limited use for complicated tasks.

Terminal tools: Ghostty + Herdr

Why Ghostty instead of the stock terminal on the Spark

  • It stays smooth under agent load. GPU-accelerated rendering keeps long agent logs, diffs, and test output scrolling cleanly — the stock terminal on the Spark stutters and lags on exactly the output we produce all day.
  • It feels instant. Native, low-latency input — when you live in the terminal for hours, the difference is not subtle.
  • Splits and tabs are built in, with sane defaults and excellent font rendering — no extra tooling needed before you are productive.
  • One config, every machine. The same dotfiles drive Ghostty on the Mac and on the Spark, so the whole team shares one terminal setup.
  • It is Herdr’s home. Herdr builds on the Ghostty VT engine and runs best inside Ghostty — choosing the terminal chooses the control room.

Herdr — one terminal for the whole herd

Herdr is a terminal-native agent multiplexer: workspaces, tabs, and panes with tmux-style persistence, plus something tmux never had — semantic agent state. Every pane shows whether its agent is working, blocked, or done, so running several agents in parallel stops being guesswork. It stays in the terminal we already use; no Electron dashboard, no separate environment.

ghostty.org ↗herdr.dev ↗

Matt Pocock Skills

Matt Pocock — the TypeScript educator behind Total TypeScript — open-sourced his agent skills collection: small, readable skills for real engineering, not vibe coding. Test-driven development, bug diagnosis, domain modeling, interface design, code review — each skill encodes a disciplined workflow any agent can follow. Many of the skills installed in our own environment trace directly back to this collection, and it is the fastest way for a CoE team to raise its agent game.

GitHub — github.com/mattpocock/skills ↗YouTube — Matt Pocock’s channel ↗

The model journey — and why Kimi now

  1. The last several monthsOpenAI GPT-5.5 + Anthropic Claude Opus 4.5

    The daily pair that carried the early work — reliable, capable, and the benchmark everything since is measured against.

  2. ThenGPT-5.6 Sol (Medium) + Claude Opus 4.8

    We relied heavily on this pair, with Fable in limited use for the most complicated tasks.

  3. The last two weeksKimi K2.7 + K3

    Some of the best results since beginning this journey — now the daily drivers.

Why the switch

Kimi is roughly half the price of Opus and, in our day-to-day work, performs better — though the two are close, and Opus 4.8 with GPT-5.6 Sol remains a strong pair. To be clear about the institutional reality: Anthropic or OpenAI will likely remain the gold standard at TenureOS — procurement, support, and governance all point that way, and for obvious reasons. But the Chinese open-weight models are amazing tools, and the Center of Excellence should know how to use them well.