I Built Three Versions of My AI Clone Before One Actually Worked Well

2 weeks. Three versions. The slowest one won.

I built an AI that thinks like me - answers questions the way I would, applies my frameworks, uses my voice. Claude and GPT are good, but they've never had full context around how I think. I'd spend hours re-teaching them my voice, my decision-making frameworks, my approach - every single conversation. I wanted something that just worked consistently.

Three versions in, I finally have one that works. But it's slower than the one that failed.

And I'm okay with that, because I know exactly what v4 needs to be.

Version 1: Stuff Everything In

The first version was deliberately simple. I dumped my entire 45KB voice profile and Clifton Strengths into one massive prompt. No chunking, no retrieval, just raw context.

It worked initially - it sounded like me. But as I added more content and learnings, it started hallucinating. If it got the facts right, it messed up the voice. If it nailed the voice, it missed critical details.

The frustration wasn't just that it broke. It was that I couldn't predict when or how.

The problem wasn't that v1 failed. The problem was that it couldn't scale.

Version 2: Tune the Parameters

I restructured the knowledge base. More chunks, better retrieval, explicit instructions on when to prioritise voice versus accuracy. I also added an AI QA review scan as an additional check. It got better - significantly better.

Then I tested it on parenting.

I was sitting at my desk after bedtime at almost 1am, testing v2. I asked it how to handle my daughter's tantrums. It came back with a three-phase negotiation framework - similar to ones I use for acquisition disputes. I sighed & closed the laptop.

That's when I knew parameter-tuning wasn't going to cut it.

Version 3: Rebuild the Architecture

One prompt can't hold multiple mental domains. Work, fatherhood, M&A, health - they all require different context-switching that a single model pass can't manage.

I'd built agentic workflows before using n8n, a workflow automation tool. I knew what structure would actually work. I took longer this time - reflected, mapped it out properly.

Version 3 is smarter. It context-switches cleanly. It doesn't apply business logic to my kid anymore.

But it's slower.

The relief of it finally working was immediate. The tension of accepting a trade-off I didn't want to make took longer to settle. I wanted fast AND smart. I got smart. For now, that's enough.

That's the trade-off I accepted.