AI Architecture in Practice — ArcKit at the Perth AI User Group

Last night I had the pleasure of presenting to the Perth AI User Group on a topic I’ve spent much of the past year living inside: how to get the speed of AI-assisted architecture work without giving up the rigour, traceability and audit trail that enterprise governance demands.

The talk — AI Architecture in Practice — centred on ArcKit, the open-source enterprise architecture governance harness created by Mark Craddock (tractorjuice), and how I’ve used it end-to-end on real client engagements here in Australia.

The problem: “can AI draft it?” is the wrong question

Every architect in the room recognised the starting point:

• Scattered documents, no single source of truth. Requirements, risks, decisions and designs live across SharePoint, Confluence, email and slide decks — and drift apart.
• Weeks of expert effort per artefact. Business cases, risk registers and reviews queue behind delivery; the expert-drafting layer is what everyone waits for.
• No answer when the auditor asks “why?” Nothing links a decision back to a requirement, a risk, or a business driver. The audit trail gets reconstructed from memory.
• Raw GenAI makes it worse, not better. Fast, plausible drafts with no provenance, no gates and no audit trail. Speed without control is a new liability, not a productivity gain.

The framing I kept returning to all night: the question is not “can AI draft it?” — it is “can I defend it?”

What ArcKit is

ArcKit turns AI coding assistants into governed architecture engines. Instead of Word documents and stale Visio diagrams, every artefact — principles, business cases, requirements, risks, decision records, design reviews — is generated by slash commands, written as version-controlled Markdown, and stamped with provenance and traceability.

Four things make it different from just prompting a chatbot:

1. Open source. MIT-licensed, with docs, playbooks and 40+ public example projects at arckit.org.
2. It runs inside your AI assistant. A plugin for Claude Code (the premier experience), plus Codex CLI, Gemini CLI, GitHub Copilot and OpenCode. Slash commands, not a SaaS platform.
3. Markdown + Git, not documents. Artefacts live in the repo alongside code — versioned, diffable, reviewable via pull request. Governance as code.
4. Governed, not just generated. A harness intercepts every write: gates block unsafe changes, observers stamp provenance, and traceability links artefacts end-to-end. The AI never writes directly to the workspace.

That last point is why it’s called a harness. Every tool call is intercepted, checked, enriched and stamped — so speed never comes at the cost of control.

The Australian overlays

Because most of my work sits in Australian regulated sectors, I maintain the arckit-au overlay — community-contributed commands anchored on federal frameworks. Compliance becomes a slash command you run in week one, not a finding you receive in the final month:

Plus sector packs for critical infrastructure (SOCI/CIRMP), OT security and energy (AESCSF). Install everything with:

claude plugin install arckit arckit-au

The live demo

A worked example of an architecture engagement, governed end to end by ArcKit rather than hand-written.

The whole demo workspace is public. Clone it and follow along:

👉 github.com/SmartTechVentures-au/arckit-demo

Two real engagements, same harness

The heart of the talk was two anonymised client examples showing ArcKit under genuinely different conditions.

Client Example 1 — community services provider (NDIS & aged care). Four concurrent mission-critical platform procurements under hard regulatory deadlines. The programme produced 96 governed artefacts across 5 project spaces, including 38 architecture decision records and 27 structured vendor profiles. A 000-global space held org-wide principles, business case and risks; each procurement ran as a numbered project inheriting from it. Structured vendor profiles plus requirements traceability gave the evaluation an auditable evidence base — decisions defensible to the board and to regulators.

Client Example 2 — ASX-listed industrial technology company. Governance for a new enterprise data platform, with a very different security posture: every artefact carries a PSPF-style classification marking (UNOFFICIAL → PROTECTED) enforced through profile configuration. Notably, this engagement ran on OpenAI Codex CLI while Example 1 used Claude Code — same artefacts, same governance, different AI harness. Assistant-agnostic in practice, not just in the brochure. The full artefact chain — PLAN → STKE → REQ → RISK → DATA → SOBC → HLD → HLDR → ADRs → AUPIA → TRAC → CONF — landed as 30+ governed, versioned, traceable artefacts.

What held across both engagements

Six patterns I’d stand behind for anyone starting out:

1. The architect’s job changes, not disappears. AI drafts to templates; the architect interrogates, decides and signs off. Judgement, stakeholder context and trade-off thinking become the whole job — formatting effort drops to zero.
2. Start with principles and a global space. Org-wide principles plus a 000-global project made every subsequent initiative faster and more consistent. It’s the highest-leverage first week you can spend.
3. Traceability changes conversations. When every requirement links to a driver and every decision to a requirement, procurement disputes and audit questions get resolved from the artefacts — not from memory.
4. Overlays make compliance a starting point. Running the AU privacy, cyber and AI-assurance commands early surfaced obligations while the architecture could still absorb them cheaply.
5. Git-native governance meets teams where they are. Pull-request review brought engineers into governance; publishing to DevOps Wiki, Confluence and GitHub Pages brought governance to executives.
6. AI output still needs qualified owners. Community overlays are accelerators, not opinions of record — security, privacy and legal owners review before reliance. The harness makes that review easy, not optional.

Getting started

1. Explore — arckit.org has guides, playbooks and 40+ public example projects across health, AI, cloud and data.
2. Install — claude plugin install arckit arckit-au. Free, open source.
3. Pilot — pick one live initiative, run /arckit:init, follow the navigator, and put the eight essential artefacts in Git within a fortnight.

Thanks to the Perth AI User Group for hosting, and to everyone who stayed for the wine, pizza and the long tail of questions afterwards — the discussion on assistant-agnostic governance and the role of the human reviewer was the best part of the night.

If you’re wrestling with AI governance on a real programme and want to talk it through, you’ll find me at mckelt.com, smarttechventures.au or axurcio.com.

ArcKit is MIT-licensed open source by Mark Craddock (tractorjuice) — github.com/tractorjuice/arc-kit. Community AU overlays should be reviewed with qualified security, privacy and legal owners before reliance.