Build a Digital Twin, Not a Digital Doll

For a while my AI setup had a style guide that described how I sign off my weekly note to my team. It listed four sign-offs. They sounded like me. They were friendly, a bit Australian, a bit warm.

Then I gave the setup all fifteen editions I had actually sent and asked it to check. None of the four sign-offs appeared in any of them. Not once. What I actually write, most weeks, is a variation of "Have a super weekend, be safe, be kind, stay awesome." About half the time I do not sign off at all. I end on a question.

The style guide was a description of who I thought I was. The fifteen editions were a record of what I did. The setup had been learning from the description.

That is the difference between a digital doll and a digital twin.

Doll versus twin

A digital doll is built from a description. A persona prompt, a list of traits, a paragraph about your values. It can sound like you. It will confidently produce things you would never say, because nothing ties it to what you have actually said.

A digital twin is built from a record. The things you sent, the decisions you made, the notes you wrote at the time, each with a date and a source. It can still be wrong. But when it is wrong, you can see why, and you can fix the record.

The rule I now run on is simple: the record outranks the description. When my style guide and my sent editions disagree, the editions win and the guide gets corrected. That rule is written into the guide itself.

The quickest way to tell which one you have is to ask your AI assistant one fact about how you work, then ask where it found it. A twin points to something: a document, an email, a date. A doll improvises a plausible answer and a vague reason. The second answer tells you much more about your setup than the first.

The model is only the mirror

I heard Peter Tonagh speak about building a digital twin and it stuck with me, partly because I had been quietly building my own for more than two years. The line I took from it and have repeated since is that the model is only the mirror. It reflects what you feed it.

Most people are about to get a workplace AI that learns from their emails and documents. So the question I asked my team was this: if your AI only knew you from this week's emails, would you be proud of what it learned?

For most of us, the honest answer is not quite. This week's emails are rushed, reactive and missing the reasons behind the decisions. That is what the mirror will reflect unless we give it something better.

Own the record, rent the head

If the record is the twin, where the record lives matters more than which model reads it.

Every few weeks someone posts an AI setup that looks finished. Agents with job titles. A memory system with a clever name. A dashboard that glows. The replies fill up with people saying they are moving everything across this weekend. I feel that pull every time.

The distinction I use now: steal the pattern, do not adopt the head. A pattern is how a setup works: how it captures, how it retrieves, where a human still has to click before anything leaves. Patterns travel. A head is what the setup happens to run on: a particular chat product holding the memory, a persona prompt, someone else's idea of who you are. When you adopt a head, you inherit the author's memory, taste and mess, and your twin ends up living inside their product.

So my files are the archive. Tools can read, draft and propose. None of them gets to become the place my work lives. Today all of my models can tap into the same vault of files: Grok, Claude from Anthropic, and ChatGPT. I have retired many tools along the way and never lost my core memory, because I put a hard focus on setting that memory up and putting good governance around it. The files and the contracts stayed. The tools rotated. That is the point.

The Friday test

The test I run on any tool that wants to hold my record is short. If this tool vanished on Friday, could I rerun Monday's work from my own files?

In practice that breaks into a few questions:

• Can I export everything, including the context and not just the outputs?
• Is the memory in files I control, or only in a thread I cannot reach?
• Could a different model follow the same instructions and get close to the same result?
• If I stopped paying tomorrow, what would I lose that I could not rebuild?

A tool that fails the Friday test can still be useful. It just does not get to be the archive.

The same question is coming at work. Vendors will offer a copilot bound to their platform, holding your organisation's memory. That can be fine as labour. It is risky as the system of record. Ask whether the memory exports cleanly and whether you could swap the model underneath without rewriting how the work runs. Some of what they offer will be a good idea, but not our path.

Repetition beats inspiration

The better record does not come from one brilliant prompt. It comes from ordinary work written down, repeatedly.

My weekly note to my team is the best source my twin has, not because any one edition is special, but because there are fifteen of them, written the same way, about real work. Speed without a standard just produces average faster. The standard comes from repetition.

The same goes for how I brief any AI on a piece of work. Before I ask for anything that matters, I write five lines: the outcome I want, who feels it, what good sounds like, what must never happen, and what done means. It takes two minutes. Over time, those five lines become part of the record too.

Guards

A twin that sounds like you is powerful, which is why it needs guards.

You own the send. Always. The twin drafts. It does not send, post or publish in my name. Anything it writes is marked as model-authored until I have sent it myself. Only then does it count as my voice in the record.

Lived, not modelled. The twin can tell me what I have said. It cannot tell me what I think now. If I have changed my mind, the record needs a new entry, not a clever inference.

Cite, or say you cannot. If the twin cannot point to where it learned something, it should say so plainly rather than fill the gap.

Honest limits

My twin is not finished. Parts of my own voice guide were wrong until recently, and I only found out because the record was there to check against. A doll would never have told me.

If your AI assistant disagreed with your own description of yourself, which one would you trust, and do you own the record to settle it?