Hey, it's been a while; I'm on a sabbatical from work. so writing is not always top of mind. It was supposed to be my time to finally ship all those side projects I'd been dreaming about.
Turns out, giving me unlimited time to build things just gave me unlimited time to rebuild the same things over and over. I spent almost all of my free time building features that don't matter. Authentication systems for apps only I use. Realtime for data that updates monthly. Theme customisation wizards when I was already happy with my theme.
Last month has been peak absurdity. A simple book writing idea morphed into a multi-tenant SaaS with realtime collaboration and offline support. Zero users. Three architecture rewrites where I pivoted between web, electron, and a native macOS app. A month of "free time" wasted.
The worst part is that AI made it easier. I could generate a solid first pass of an app in minutes. Data access, collaborative editing, on-point UI components - all instantly generated, all completely unnecessary.
Claude Code wrote perfect code for the wrong product. Windsurf helped me polish features nobody wanted. We've made it terrifically easy to build the wrong things.
With unlimited hours, I just spent longer debating with myself. Should it be a CLI or GUI? Web or native? I had entire days to sift through every possible architecture.
My calendar was empty but my mind was in back-to-back meetings with myself.
Why user research wasn't the answer
You might think the solution is simple: talk to real users. I've done that. The problem with traditional user research is that users anchor on what they know. Ask them about a writing tool, they'll describe Microsoft Word with better spell check. Ask about story organisation, they want Scrivener but faster.
Most can't imagine what's possible with AI. Few know they could have a tool that adapts to their writing style, or automatically maintains character consistency, or generates chapter outlines from scattered notes. Most ask for faster horses, not cars.
But building AI-first products without any user input is how you end up with my sabbatical disasters - technically impressive solutions that nobody wants.
After weeks of spiralling, I had a thought: what if I created digital twins of my users? Not just any personas - AI representations who understand both current pain points AND what's possible with AI?
Building users before building products
I created a Claude project with detailed personas of my "target users":
- Story Writer Sam: Writes short fiction on weekends. AI-curious.
- Indie Author Maya: Self-publishes 4 books a year. Willing to try new tools but needs quick wins.
- Publishing Pro Emma: Manages multiple series at a small press. (She doesn't exist in my market yet.)
- Hobbyist Hannah: Writes fan-fiction for fun. Zero patience for complexity.
I gave each persona knowledge of AI capabilities but also realistic constraints. Sam knows AI can help with plot consistency but worries about losing creative control. Maya understands automation possibilities but has established workflows. Hannah has seen AI demos but still writes on her phone.
Then, before building anything, I pitched my ideas. Each persona responded from their perspective:
Me: "AI-powered collaborative editing with realtime suggestions!" Sam: "I write alone. I already turn off grammar suggestions." Maya: "I need to maintain my voice/tone - won't AI turn it to mud?"
Me: "It uses AI to automatically structure your manuscript!" Hannah: "I don't want AI organising my story!" Sam: "Could it track my plot threads and character arcs instead?"
Me: "Native app with local AI for maximum performance?" Everyone: "We write across devices. Make it web-based & responsive"
The conversations revealed something crucial: even AI-aware users don't want AI everything. They want AI to solve specific problems without adding complexity.
The balance of innovation and adoption
Having these digital twins taught me something traditional user research couldn't: how to find the sweet spot between AI capabilities and user readiness.
Real users would have asked for a better text editor. My digital twins helped me discover they actually needed AI that tracks character consistency across chapters - something they didn't know was possible but immediately recognised as valuable when presented.
The pattern became clear. I wasn't failing because I built too much or too little. I was failing because I built for imaginary users who conveniently wanted features I thought were cool.
My "Story Writer Sam" was actually "Story Writer Who Secretly Wants Enterprise Publishing Software Sam." My "Indie Author Maya" was "Indie Author With Unlimited Budget and Time Maya."
Real features for realistic personas
Now, before opening code, I run a user panel with Claude. But it's not about asking what features they want - it's about understanding what problems they'd pay to solve and what complexity they'll tolerate.
Here's how a conversation went for a book cover tool I built to round out my sabbatical:
Sam: "I spend $50-100 on Fiverr for covers. I'd pay $20 for instant generation IF it doesn't look AI-generated." Hannah: "I make my own in Canva. Pass." Maya: "Can it do series branding? Same style across multiple books?" Emma: "Oh, actually, that's exactly what I need too!"
Three paying users out of four. But they had specific needs: simple, fast, maintains series consistency, doesn't look AI-generated. No collaboration. No teams. No fancy features.
After 105 revisions of an Artifact within Claude Desktop (yes, really), I had something they may actually use. I moved it to a little React app and the whole thing took an afternoon. It solves one problem well instead of ten problems poorly.
The users who don't exist (yet)
Digital twins revealed something else: I kept building for Publishing Pro Emma, who doesn't exist in my market yet. She's the user I hope to have in three years. But Sam and Maya exist today and will pay today.
This approach - digital twins who understand AI but maintain realistic scepticism - solves the innovation paradox. They push you beyond what current users imagine while keeping you grounded in what they'll actually adopt.
Traditional user research would have told me to build a better Scrivener. Building without users gave me a multi-tenant platform nobody wanted. Digital twins helped me find the middle: an AI-first tool that solves a real problem in a way users are ready to adopt.
The real revolution
This process made me reflect on the current state of AI in software development. We've made code generation instant while leaving all the decision-making overhead untouched.
The bottleneck was never writing code.
In the case of user research, the goal was to bridge the gap between what users say they want (faster horses) and what's actually possible (cars they're ready to drive). I daresay there is scope for a lot of improvement across all areas of the software development lifecycle too. These unlocks are far more interesting than speeding up the code churn!
So maybe the next revolution isn't The AI Engineer. Instead it's AI that helps us hear what our users actually want - even when they can't articulate it yet.
Now if you'll excuse me, Story Writer Sam wants to discuss a feature for my book cover generator. (Sam will probably say no. That's why Sam is valuable.)