About

Reflection

About

This site brings the stages of the project together as one piece, from the first question about rain to the developing precinct. It is a draft. What is unfinished is named on each page and gathered at the end of this page.

How I use AI in design practice

Position

I use AI as a production team that works under my direction, while the design decisions stay with me. Across Common Ground, AI researched, scripted, drew, rendered and built the website. I set the brief, made every design call and decided what was good enough to show.

I’ve discovered that the processes behind architectural design are where I find some of my greatest satisfaction. I enjoy understanding how things work, identifying what could be improved and developing better ways to achieve the outcome. That doesn’t make the architectural outcomes less important to me. It means I care as much about the system that produces the work as about the work itself.

The detailed method, with the rules, skills, and review records, is on Approach. This statement explains what I learned and how I would bring it into practice.

Protocols

Every tool works under a short set of written rules, and most were added because something went wrong: authorship, no silent omission or transformation, the repeat gate, approval gates and never deleting. Each rule, and the mistake behind it, is set out on Approach.

The rules live in my Obsidian vault, alongside handovers written at the end of each session. A new session starts by reading them, so the work picks up where it left off instead of starting from memory.

The vault works because I keep it strict. Each kind of record has one place, files are named consistently, there are no unnecessary folders, nothing is deleted unless it is saved elsewhere, and every session ends with a handover. AI is only as reliable as the context it is given, and a disciplined vault gives it the right context every time.

I started scheduling with AI at the site reading stage. For the site reading and the canopy, the schedule was written into the vault and approved by me before the work began. The routine around it, checking what was left and warning me as each hard stop approached, was never written as a rule, so it depended on each chat carrying it out. Next time I would make it a rule in the vault from the start.

Benefits

The biggest benefit was integrating AI with the software I already use. Connected to Rhino, Grasshopper, Photoshop, Illustrator and WordPress, it automated the mechanical work and let me run several tasks at once. It also kept the project knowledge organised and helped me manage my time.

  • Depth within the time. One designer could test canopy structure in Karamba3D, work out a site water balance and draw a 1:50 section through a flood-ready building, all within a short course.
  • Production from the model. The 1:50 section was generated from the Rhino model’s cut data, so the drawing matches the model. My time went into what the drawing should show, not into tracing lines.
  • Continuity. Handovers and the vault meant that each session started from my last decisions, not from a blank page.
  • Building the workflow itself. I enjoyed continually developing the workflows: approval gates, omission alerts and reusable skills. My Obsidian vault became central to this.

Managing Risk

AI is confident whether it is right or wrong, so I treat its output as unverified until it is checked.

  • Sources. Every site fact on a drawing has an entry in a verification register, checked against the primary source, such as Brisbane City Plan overlays and council flood data. An empty dataset is recorded as empty, never read as “no constraint”.
  • A second system as a critic. ChatGPT reviews outside the production chain, so the check comes from a different system than the one that did the work.
  • Honest limits. Structural and hydraulic work is presented as the designer’s proposal, subject to engineering review, and never as engineered. Open items stay visible in the documentation rather than being smoothed over.
  • Questioning the output. When the 1:50 section showed a flat collector pipe between tanks, I questioned how the water would move sideways. That question changed the design to a balance pipe that always runs full.
  • Copyright and privacy. Licensed standards are used for checking only and are never shared or published, and other people’s work stays out of my records.

Generation, Iteration and Curation

I generate widely at the start and narrow quickly. Midjourney gave the early atmosphere studies, and I moved the chosen images into ChatGPT for precise, targeted edits. From there, the work moved into the model, where it could be measured.

Iteration happens in small, inspected steps. The 1:50 section went through more than twenty versions in a day. Each one changed a single thing I asked for, such as moving a seated figure clear of the cut or lightening the shade under the canopy.

Curation is where my judgement shows most. I set rules for what makes the cut: nothing may sit in front of cut material, and shadow means “under cover” and nothing else. When a rule didn’t hold, I removed the element.

The hardest lesson was about renders. AI images gave me the atmosphere I wanted, but they drifted from the model: missing steps, invented balconies, a shifted camera. I learned that AI cannot add accuracy. Accurate, atmospheric renders come from an accurate, detailed model in a dedicated renderer, and AI can only fill the gaps the model leaves. That is why I paused rendering until the model was reviewed, and why I moved to Lumion. In Lumion, a clash between the steel and the collars showed up in every view, so the model still needs that fix before the final render.

Tools

Each tool has one job, and I added a tool only where it did something the others could not. Claude worked inside Rhino and Grasshopper, Nano Banana, Photoshop, Illustrator, InDesign and WordPress through MCP connections, and read and wrote the vault files directly. Karamba3D ran inside Grasshopper through the same connection, and Lumion took the live model through LiveSync from Rhino for review with materials, where it exposed the steel-through-collar clash. The final render follows once the model is corrected. The full chain, stage by stage, is on Approach.

Grammarly runs on everything I write on my computer and suggests corrections and rewordings as I type. I accept or reject each one, and the writing stays mine.

I gave each project its own connectors, so this course’s tools only reach this course’s files. I also wrote three reusable skills: a concept critique, a site reading harness and a model and documentation review.

Outlook

I think practice is moving towards designers who direct AI production rather than doing every step by hand. The value will sit in judgement, in the quality of the model and in the rules that keep the output honest. A practice that writes those rules down will get consistent work from AI. A practice that doesn’t will get confident mistakes.

This course showed me that my strengths sit between design, computational work and the systems that connect them. I would bring a working method to a practice: written rules, reusable skills, a verification habit and a record of every decision.

Next, I want to keep developing methods for accurate renders, so that atmosphere and accuracy come from the same model. I also want to take the documentation workflow further into Revit through Rhino.Inside, and keep refining the skills so they work on projects beyond this one.

What is left open

Each page lists its own open items. This is the index, with the items that belong to no single page.

  • The Proposal. A second fire stair, daylight to the deeper floors, the front columns, rain off the projecting apartments, canopy lighting, the community uses, and the model corrections before the final render. The Proposal
  • Testing the Canopy. The exploded axonometric redraw, spot level 6, the ETFE cushion foils, the trunk member spacing, the structure for an engineer’s review and rain noise on the ETFE. Testing the Canopy
  • Storage and drawdown. The collection chambers and the balance pipe between them are drawn in the 1:50 section. Sizing the storage in the existing silos, the overflow, treatment and drawdown are still to be designed, with a hydraulic engineer.
  • The Smart Farm building. Plans and sections of the farm are not yet drawn, and the choice between daylight and artificial light for growing is still open.
  • The farm behind glass. Whether the growing rooms are seen through a glazed wall is still open, because the concept asks that the water itself not become spectacle.
  • Water quality. The canopy skin study set three conditions for the stored water; treatment is allowed for at each downpipe’s handover and not yet designed.
  • A reading of the river edge. A reading of 61 Montague Road against the flood and cadastre plates of drawing D01 is still to be written.
  • Documentation software. The project was modelled and drawn in Rhino and Grasshopper, not Revit.
About this site

Who it is for

A prospective employer: an architecture practice reviewing my portfolio.

Three real examples studied

ExampleLead withHow muchLeave out
Bespoke Careers: four employers review three real portfolios, including Arup ArchitectureThe strongest, most relevant work first; a summary page with each project and your role8 to 15 pages; reviewers spend seconds and most stop within the first few pagesSoftware icons (write names as text); mixed page sizes and orientations; transfer links
Bespoke Careers: design portfolio tips from four hiring employers, including a Brisbane practiceThe strongest work, straight awayProcess on one page is welcome; the portfolio must read without the candidate there to explain itContents pages, links and QR codes
DesignInc: architectural graduate job advertisement (Dec 2024)A succinct portfolioNo page limit statedNot stated; asks for the methods and software used to be shown with the work

How it is arranged

An AI brainstorm set out three ways to arrange the same five components, each for a different audience. I kept a mix of the first two.

Opening kept

A Practice Employer

Leads with drawings, the role and the software as text; about 12 pages, most of them the resolved canopy.

Order kept

A Jury Assessing AI Practice

The five stages in the order they were made, each showing the decision, what the AI did and my call.

Not kept

The Client, Common Ground Food Lab

Leads with atmosphere: the vignettes and the sound of rain, with the AI method reduced to one line.

I chose this arrangement because I want the website to become my professional portfolio, so it needs to show my architectural design capabilities straight away. The current proposal comes first, followed by the development stages in order, because I also want to demonstrate how I work with AI, how the design evolved and the skills I’ve developed throughout the course.

Brief to myself for the final project

Written before the site was restructured on 26 Oct 2026, and kept as written. Some page names have changed since.

My final project will be presented as a website aimed primarily at prospective employers in architectural practices. I want it to demonstrate my architectural design capabilities alongside my ability to integrate AI into the design process, from early concepts through to modelling, documentation and review.

The website will lead with the strongest architectural work, including the more resolved canopy design, before taking the viewer through the stages of the project: Rain, Ground, Concept, Capture, The precinct and Check. I want to show not just the outcomes, but how the design developed, the decisions I made and the ways AI supported the process. The website will also provide a foundation for my ongoing professional portfolio.

My main priorities before the final submission are to redraw the exploded axonometric and improve the Capture concept image. I want the axonometric to communicate the canopy assembly more clearly. For Capture, I’ll return to one of the stronger early concept images because the current image doesn’t accurately represent how I envision the funnel system working or how it integrates with public life. The glazed growing rooms and the way water is shown on the roof also don’t reflect my design intentions. I want the revised image to better communicate the atmosphere and spatial experience I intended.

Separately, I’ll use a visualisation developed from the Rhino model as the main website hero to represent the more resolved canopy design.

I’ll also correct known documentation errors, including spot level 6, and review the remaining gaps against the time available. Additional drawings and technical refinements will be considered where they genuinely improve the presentation, rather than simply adding more material.

Ultimately, I want the project to demonstrate both my architectural thinking and the practical ways I’ve learned to use AI while maintaining control over the design direction and decisions.