Adventures in AI Prototyping

Like the whole industry, I've been experimenting and learning with AI. From chatbot assistants like ChatGPT, Claude, and Gemini, to prototyping tools like Lovable and Figma Make, to production-grade programming tools like Cursor, Antigravity, and Claude Code.

On the one hand, it's remarkable technology. But on the other, there's a whole lot of externalities to think about (energy usage, theft of intellectual property, hallucinations, and rewiring of human brain circuits to name just a few).

I think it's highly likely that we're in a bubble and the resulting crash is going to cause a lot of pain. There's just no way for the industry to generate the returns needed on the investment to date (and let's be real, AGI ain't happening anytime soon). But just because the stock is overvalued, doesn't mean the technology is worthless. We've seen this type of hyper-inflated story before — the wreckage of the dot com crash was real, but it also provided the kind of reset needed for a prolonged stage of economic growth in tech.

Setting the Wrong Goals

2025 featured a lot of conversations about AI as a productivity enhancement. And, like many, I've been skeptical about how attainable these goals actually were.

With the data that's starting to come in, it's starting to look like AI isn't going to be the productivity tool it was hyped up to be, and many of the original claims have since been busted.

So let's set marginal productivity gains to the side. In 2026, I hope that instead of talking about productivity, we talk about enablement.

I know that in my case, AI hasn't made me more productive — but it has enabled me to do things that I couldn't before by lowering the barriers of time and expertise. If I can build a functioning prototype on my own (instead of with an engineer), in a day or two (instead of a month), then that fundamentally changes my approach to exploring and testing new ideas.

And if we extend that kind of thinking to the organization level, there will be problem spaces where companies previously chose not to compete, now open for renewed competition. Leaders would be wise to check their assumptions, because the old rules of the game no longer apply.

Figma Wrapped

One example of something I've been able to build recently is a Figma plugin. Now, it should be stated that I have never built a Figma plugin. In fact, I've barely skimmed the documentation so this is a tech stack that I am decidedly not familiar with.

But, after trying my hand at a few Figma Make prototypes, I was looking for a project to step into a real programming tool. Google had just launched Antigravity and the model access was free, so it seemed like a good time to try something new. And truthfully, I was feeling a bit anxious after hearing Ryan Smith and David Gratton speak about their agentic coding workflows at ProductBC back in October.

So in a single evening, I mapped out some specifications for a "Spotify Wrapped"-style Figma plugin and got to work. I used Claude to scan the developer documentation and query about what was feasible, then broke the concept down into approximately ten key steps.

• A basic hello world
• File key and access token setup
• Basic plugin API data retrieval (number of artboards, layers, etc)
• Basic REST API data retrieval (version control, comments, etc)
• Editor and commenter leaderboards
• Most used component leaderboard
• Designer style calculation (autolayout vs freeform)
• Favourite fonts and colours
• Export to PNG functionality

Each of these steps required some back-and-forth, debugging, and re-prompting (some more than others), but chunking it out in this fashion was helpful to keep me (and the AI) focused on one task at a time. I suppose it's analogous to breaking a large feature like an Epic down into smaller tickets or Stories for sprint planning.

Would I ever ship this code? Of course not. But when your job is to bring new ideas to life, it's hard not to look at these tools with at least a little bit of amazement. This is a real, working proof-of-concept — it's far from perfect and needs more soul and storytelling to be sure — but it's pretty impressive for a single evening of work.

Bring Your Own BS Detector

Now I thought I'd end this article with a note of caution.

We’re still all just starting to learn how to use AI, and it can do a lot of incredible, almost magical things. But if you’re familiar with the Dunning-Kruger effect, there’s a point on the journey to expertise where competence is still low, and confidence is sky high. This is the point some of us (and many influencers) are at in the AI hype cycle, and it's actually kind of dangerous.

We need to temper our enthusiasm for this new tool with an understanding of its strengths and its weaknesses. And in my view, what this really comes down to is this: as the human operator of a large language model, it’s your responsibility to bring the BS detector.

AI is probabilistic. It aims to please rather than saying no. It doesn't actually know anything. And by now, we should all be familiar with AI’s tendency to hallucinate with unearned confidence.

So yes, you should continue to experiment with AI. I think the technology can be a powerful enabler. But please do so responsibly.

I'd love to hear your thoughts. What has AI enabled for you that was previously impossible (or near impossible?)