Vision
Why a Reasoning Graph, Not a Prompt
The persona prompt
The simplest way to build an AI version of someone is to describe them in a prompt. Bio, writing samples, a few documents, "answer as this person" at the top.
More careful versions split the documents into chunks, store them in a vector database and retrieve the relevant ones for each question. The principle is the same: turn a person into text and let the model work out the rest.
We should be fair about this. It takes an afternoon, it costs very little, and for many jobs it is enough. If you want an assistant that answers questions about your published material in roughly your tone, build a custom GPT and stop there.
It stops being enough when someone needs your judgement.
Where it stops
Take a question a client might put to an adviser's twin: "Our biggest customer wants a 15% cut to renew. Do we agree?"
A persona prompt with retrieval will find your article on pricing, and perhaps a proposal where you held firm. It will write a fluent reply in your voice. What it does not have is any record of how you weigh this: that you hold the rate when the objection is budget, that you trade price for term but never for volume promises, that you treat the customer differently if they are over a third of revenue.
It knows what you have said. It does not know how you decide.
Three specific limits follow.
No conditions. A description of you is flat. "Firm on pricing" and "flexible with good clients" are both true of most people, and a prompt cannot say which applies when.
No explanation. Ask why it answered that way and it produces a rationale after the fact. There is nothing behind the answer to point to.
No memory of being wrong. You correct it in one conversation. The next conversation starts from the same prompt. To make a correction last you edit the prompt by hand, and the prompt grows into a list of patches.
What Imora stores
Imora keeps a reasoning graph for each person. It has three parts.
Evidence. Your answers to real scenarios, your corrections, things you tell your own twin about how you would decide, and views stated in documents you wrote. Kept verbatim and dated.
Principles. What Imora distils from the evidence: rules in your terms, with the conditions and exceptions attached, each linked back to the words it came from. You can read them, and you should.
Links. Related principles connect. Where two pull against each other, such as "protect the relationship" and "never work unpaid", the tension stays visible and is not averaged away.
The language models that write the replies are Anthropic's Claude, the same family you could use in a Claude Project. We are not claiming a cleverer model. What is Imora's own is the graph and how it retrieves your decisions.
The difference in practice
Ask the twin the renewal question. Imora works out what kind of decision it is and what is at stake, retrieves the principles and evidence that fit, and the twin answers from those.
The reply opens with the twin's read: hold the rate, offer a two-year term at a smaller reduction, and check how dependent you are on this customer before the call. "Why this answer" shows the principles used, the scenario answers they came from, and how confident the twin could be.
If you disagree, mark it "Not what I'd say" and write your version. That correction outranks everything else in the graph and applies from then on. "Sounds like me" confirms a reply that was right.
And if there is nothing to go on? A shared twin says it has no recorded view. It can offer an approach, clearly framed as an approach and not your position, and the question lands in your inbox. A persona prompt in the same spot will write something confident under your name.
Who gets to change it
A prompt-based assistant can often be steered by whoever is talking to it. With a twin that speaks for you, that is a risk.
On Imora only the owner teaches the twin. Nothing a visitor says changes how it thinks. Visitors' questions reach you. Their opinions do not reach your graph.
One graph, several audiences
A prompt has to be rewritten for each audience. The graph does not. An Imprint is the model of you; twins are versions of it for different audiences. A client twin and a team twin set their own focus and tone and reason from the same evidence. We make that case in build once, deploy many.
The bottom line
You can build an AI version of yourself with a prompt, and sometimes you should.
If people will act on its answers as your judgement, you need three things a prompt does not give you: conditions, explanations and corrections that last. Getting started takes about 10 minutes and five real scenarios. The product page shows the steps, and the comparison page puts Imora beside custom GPTs and clone tools, including where they are ahead.