Congratulations! You're An Architect.

My personal journey into the future of engineering

• Team Lead

• Manager

• Product Owner

• Conductor

• Director

• Orchestrator

Any one of those words could have replaced “Architect” in this post title. Of course, I’m talking about the role of SWEs in this new era that has flickered into existence over the past few years.

I’m still surprised that so many are resistant to what has for years seemed to me to be an inevitability: AI would eventually do the actual coding, humans would direct them. I think part of the problem is that people’s ideas about AI capabilities are outdated. It’s not their fault; things are progressing quickly. The models keep improving, and so does the tooling. Between the two, many (engineers and non-engineers alike), would be blown away by what is actually possible now.

Hello, GPT-3

I signed up for ChatGPT on December 3, 2022. I think I became a paying customer within the same week. I knew what I was seeing and embraced it immediately, because I’d been waiting for it since I first saw GPT-2 spitting out barely coherent nonsense. The trajectory was clear to me, but it became real the first time I put ChatGPT to the test and asked it to write a script I’d already implemented. I didn’t show it the script, just described what I wanted it to do. It wrote a better version and in only ~10 seconds. It had taken me ~45 minutes. My job was never going to be the same again. The same goes for so many others. Perhaps all others?

Hello, Claude

AI has come a long way since then. But if ChatGPT was the start of AI’s integration into my work, finding Claude was what gave me a glimpse of the bigger picture. My GPT workflow was long conversations going back and forth, correcting its understanding, copy and pasting code, running it, maybe it worked maybe it didn’t, going back, rinse, repeat.

Then Anthropic dropped artefacts in Claude. How many of you remember that moment? Artefacts + Projects in Claude was an accelerator. Suddenly I could have all relevant context for a project in one place, and even have Claude produce files and easily add them to the same project. The ergonomics of working with Claude were so much better than working with ChatGPT. On top of that, Claude’s “personality” and “demeanour” just worked so much better for me. I put those words in quotes because it still feels a bit strange using them to talk about AI. It is what it is.

Accumulating Subscriptions

Despite discovering Claude, I didn’t drop ChatGPT, because their capabilities complemented each other. If there was something Claude struggled with, I’d try ChatGPT, and vice versa. Claude also had significantly lower usage limits than ChatGPT. It still does. But Claude was that good that I was willing to pay the same price for lower limits. Whenever I’d hit those limits, though, I’d fall back to ChatGPT.

According to my billing history, I started paying for Claude in June 2024. Since then I’ve accumulated more subscriptions:

• Gemini

• Z.AI

• Manus (now cancelled)

I have 4 active at the moment. I’ve no doubt I’ll either accumulate more this year, or upgrade to more expensive plans. I just upgraded to Claude’s 5x Max plan, for reasons I’ll get to later. With AI so deeply integrated into my work (and my life), and the likely advancements we’ll see this year and onwards, my AI use will undoubtedly increase. Probably significantly so.

Hello, Cursor

Some of you would have heard of Devin, a platform that promised to build applications completely autonomously based on the user’s request. The idea was appealing because I have a lot of app ideas, but the price was too high. I also didn’t like not being involved in the process. AI has a habit of spewing out junk code: possibly working but poorly structured, bloated and brittle. Luckily there was a better suited alternative.

Cursor is an Integrated Development Environment (IDE) with agentic AI capabilities deeply integrated. In case you’re not familiar, a couple of definitions.

• IDE: An editor with features that makes it easy to navigate, understand and work with your codebase.

• Agentic: Capable of taking independent action toward a goal, not just responding to prompts.

I paid for Cursor on May 8, 2025. I’m not sure what I used it for first, but I do know that I have a conversation in a Cursor project going back to May 9th. The project is an app I built to assist my learning of the Japanese language.

Cursor allowed me to take everything I’d learned from working with ChatGPT and Claude, and apply it within a single operating environment (the IDE) while being able to switch between models. I’d flesh out a plan for a feature, provide any necessary context, then set the Claude Sonnet to work and go do something else in the meantime. I’d check in periodically and see files being created and edited, and code verified through tests. Minutes later, the AI had finished.

New level unlocked.

Hello, Ralph

November of 2025, I was feeling burned out. Work, personal issues, internal conflict, accumulating projects and ideas; it was all too much. I decided to stop working on any side projects or trying to “achieve”, and just focus on exploring, learning, and enjoying discovery with no end goal. It took the pressure off and reduced my anxiety.

Christmas came around and I started seeing “Ralph Wiggum” come up on multiple platforms. If you’re familiar with The Simpsons then you know the name. In this context it referred to a technique developed by Geoffrey Huntley.

It’s named after Ralph Wiggum because the character is relentlessly optimistic and determined. One thing you learn from working with AI is that they are incredibly optimistic. Geoffrey Huntley came up with a simple way to add the determination, by creating a simple loop.

1. You describe what you want done

2. AI attempts the work

3. AI checks if it succeeded

4. If yes, move to next task

5. If no, go back to step 2

I was intrigued but put it on the back-burner for later. “Later” was a weekend in early January. I read about the technique, then had AI create a script, task and progress files, then ran the loop. I saw the AI bring the same enthusiasm and determination to its tasks, revisiting the same task over and over if its reviews failed after each attempt. Some tasks were completed first time, others took multiple attempts, but they got there eventually.

Click.

New level unlocked. Again.

100%

ChatGPT changed the nature of my software development. Suddenly I could offload building of software to other software. It translated to only ~10% reduction in coding in the very beginning. But it also gave me access to information faster than ever, which meant the time and effort saved was actually higher.

Claude and its artefacts reduced my manual coding even further, maybe taking 30-40% of the load. With Cursor, and later tools like Claude Code, AI was writing ~80% of my code.

I’ve only been using the Ralph Loop for a couple of weeks. I’m continually learning what works, defining and refining my own flavour of the technique. But since I’ve started using it, I only review code and features. I don’t write any code.

It’s Not Slop

In the past couple of weeks I’ve built so much more with far less manual effort than I ever could have before. This is why I decided to upgrade my Claude plan—we’ve been spending far more time planning and Claude has spent far more time reviewing code.

This is my workflow.

1. Detailed Conversation: Typically with Claude. I explain what I want to build, the use case (for context), the approach(es) I’m considering, the limitations and preferences.

2. Refine Plan: Once I’m satisfied Claude has enough info, I have it produce a detailed plan with tasks broken down and testing and verification requirements included. I review and we refine.

3. Prepare Loop: I have a custom ralph-loop skill (a reusable set of instructions for AIs) that Claude uses to build the task list, progress document, guardrails, task prompt, and review prompt.

4. Run the Loop: I run the loop via a custom command line tool and peripherally monitor the loop’s progress.

5. Review/Test: Once complete I check the work. If it’s good I move forward. If it’s not, back to step one.

I’ve seen many times on Reddit and X that AI just produces slop. But while it can be the case, it doesn’t have to be. With proper context, planning, rules and guidance, good quality output is possible. It has been for many months. It varies by domain and—especially in the case of code—language, but the reality is that AI is far more capable than most people realise.

Mindset

I think a mindset shift is in order. As AI capability increases, manually writing code has diminishing returns. It can be an uncomfortable adjustment, but people need to start looking at AI as a team mate; co-worker; staff; and learn to collaborate and delegate.

Interestingly, it seems non-technical people coming to AI embrace its power more readily. Suddenly they can build a prototype when previously it was beyond their reach without hiring a developer. All for ~$20 per month on a common AI subscription.

Our jobs are changing. This isn’t just about code. It’s about all work that requires research, logic, synthesis, data entry; digital and non-digital work alike.

Architects design and plan, then others do the manual labour. Code was always a kind of manual labour. We don’t need to get our hands dirty anymore. Sure, we’ll visit the construction site, but we’re not laying any bricks. And we’re still responsible for the output, making sure the construction matches the blueprint.

Your Role

We’re in the midst of a paradigm shift. I implore you to stop thinking in terms of the old paradigm. Embrace AI fully to drastically augment your own capabilities. Make AI your collaborator; make it your co-worker. AI will happily do the work, you just have to tell it where to focus its efforts. Working with AI requires you to take on additional roles. So be a Team Lead. Be a Manager. Be a Director. Be an Orchestrator.

Be an Architect.