.png)
September 13, 2026
A better model will not sound like you, because capability and context are different problems. Every platform switch resets the relationship, and the fix is a portable record of how you actually make decisions.
For a long stretch, about eighty percent of my AI work happened in Claude.
I still tested new tools and new frontier models as they came out. That is part of being a practitioner. You can't teach the differences if you're only reading release notes. You have to use the tools, push them, and notice where each one helps you think better.
But the deep work stayed in one place. Big projects. Coding. Skills files. The reference documents that captured how I write, what I care about, and what I will send back for another round.
Then I went back to ChatGPT to help draft a newsletter.
The model had improved. I could feel it in the reasoning. Yet the first draft was strangely unfamiliar. Polished. On assignment. Not mine. The rhythm was off, the point of view was generic, and I could not find my own judgment anywhere in it.
It was almost as if ChatGPT did not know me anymore.
Then the obvious hit me. It didn't.
I had spent months teaching another platform through my edits, my examples, my Markdown files, my project history. I built a rich trail of context in one place, then walked into another room expecting the relationship to follow me.
I changed rooms without bringing my thumbprint.
So I rolled up my sleeves. I brought over the files that hold my voice, my standards, my structure, and the choices I make when there is more than one good answer. The next output moved closer.
"Cada cabeza es un mundo." (Every mind is its own world.)
AI doesn't enter that world because you typed your job title into a profile. It gets closer when you show it how you think, what you notice, what you reject, and why. And even then it only gets close.
Because the last part is not information. My mother had a word for it: sazón. Seasoning. The ingredients are available to everyone. Your hand, your memory, and your sense of what the dish needs make it yours. That is what I added after the files were in place, and that is when the work became recognizable again.
Most of us leave the part that can be written down scattered. One example lives in an old conversation. A useful correction lives in another. Our best work sits in documents the new tool has never seen. We remember the frustrating draft, but we never capture the principle behind the frustration.
Then we switch platforms and judge the new tool by its first attempt.
That first attempt is answering a different question than we think. Not "how good is this model?" How much of your thinking can it actually see?
A professional biography tells AI that you lead a business, advise clients, or run projects. Mine doesn't tell AI that I send back any draft with a dash in it, or that when the decision maker in an example turns out to be a "he" for no reason I rewrite the scene, because that detail adds nothing except an assumption, or that I call it a diagnostic and never an assessment, and I have reasons. None of that is in my job title. All of it is how you would recognize my work.
Those are judgment calls. They are also the source of the work people recognize as yours.
1. Show a model your history and the result moves, without touching the model. In the LaMP benchmark, published at ACL in 2024, researchers gave language models access to a user's own history and measured the difference across seven tasks. Results improved by a relative average of 12.2 percent with an off the shelf model, and 23.5 percent when the model was tuned on that history. Hold the size of the numbers loosely, because the tasks were structured ones like tagging and headlines rather than open writing. Hold the direction tightly. The variable they changed was not the model. It was whether the system could see the person.
2. Your own examples carry you further than you expect, and stop exactly where your voice begins. In a peer reviewed study presented at EMNLP in 2025, six frontier models were each given five samples of a real person's writing and asked to write as that person. An authorship verification classifier then judged whether the result came from the same author. For news and email writing, it said yes 95 to 97 percent of the time. For forum posts, 50 to 66 percent. For informal, blog style writing, 17 to 21 percent. More than 40,000 generations per model, across writing from more than 400 real authors.
Read those two together and you get the whole shape of the problem. Your accumulated context does real work, and it does the least work in exactly the register where people recognize you. Structured writing can be imitated from five samples. The way you sound when you are being yourself is not.
That is the gap your judgment fills. Not because AI is weak, but because what is most yours was never written down anywhere it could learn from.
Why we believe it: Every release arrives with benchmark wins and better reasoning. It makes sense to assume a more capable model asks less of us.
Reframe: Capability and personal fit solve different problems.
A frontier model arrives with broad ability. It does not arrive with your history. It has not sat in the meetings that shaped your point of view. It doesn't know which polished phrases make you cringe, which risks you will accept, or why two answers that look equal on paper feel completely different to you.
Portable context closes part of that gap. Your live judgment closes the rest.
The most useful personal AI system has three layers. The first explains who you are and what you do. The second shows how you think, through examples, decisions, and corrections. The third is you, present in the work, adding the sazón no file can fully capture.
What to do: Stop treating your AI history as something that belongs to one platform. Build a small set of files you can carry into any company approved tool: your voice, your standards, your recurring processes, and a few examples of work you are proud to put your name on. Add the reason each example works. Then keep editing. A context file is a starting point, not a substitute for attention.
Anthropic's own documentation describes how to bring your memory from another AI tool into Claude. You open that other tool and ask it to list every memory it has stored about you. Then you copy the output and paste it in. Anthropic calls the feature experimental and still in active development.
That is not a workaround someone invented. That is the documented procedure. That tells you something worth knowing: nobody is coming to move your context for you. So build it yourself, deliberately, in a format you own.
This is the work we do at FlipWork. People who go through the FlipWork Sprint spend eight weeks on it, and they leave with a Reinvention Passport: their measured FlipFactor diagnostic profile, eighteen structured skill files built from the artifacts they produced, and the agentic workflows they deployed along the way. It is portable by design. It belongs to the person and not the employer, and it uploads into whatever model they are using this year or the next one.
I am telling you that so you know the difference between the version you can build this afternoon and the version that takes a cohort to produce. A file you write about yourself captures what you already believe about yourself. A diagnostic measures what you actually are, including the gaps that confidence hides. Both are worth having. Start with the one you can do today. If you want the full version, it lives at flipwork.ai.
Start with one Markdown document named My AI Thumbprint. Markdown is plain text with light formatting, which means nearly every major AI platform can read it.
Your file answers five questions.
What do I believe about my work?
How do I decide when the answer is not obvious?
What does excellent work look and feel like to me?
What patterns do I reject, even when they are technically acceptable?
What are three examples that sound, think, or operate like me?
Keep it somewhere you control. When you test a new platform, start by giving it the same thumbprint. Now you are measuring how each tool works with your context instead of how well it guesses from a blank chat.
Run this 30-minute test:
Prompt to steal:
I am giving you a file that captures how I think, decide, communicate, and evaluate quality. Study it before you begin. First, tell me the three judgment patterns you see. Then ask me the two questions that would most improve your understanding of this task. Draft the work only after I answer. When you finish, identify the choices you made because of my context and flag any place where my judgment is still required.
1. Find one correction you keep making. Open a recent AI conversation and look for a change you had to make more than once. The tone was too formal. The recommendation ignored a stakeholder. The answer rushed toward efficiency when you would have protected trust. Write it as a standing principle and add why it matters.
2. Save decisions, not only preferences. Most personal context files are full of adjectives: warm, strategic, concise, bold. Those leave room for anything. Add one decision where you chose between two reasonable options. What you chose, what you gave up, what value guided the call. That is where your thinking becomes visible.
3. Test portability before you need it. Take the same file into a second company approved platform. Give both tools the same meaningful task. Compare where each one follows your context, where each one resists it, and where your own intervention changes the result. You will learn more from that than from any model ranking.
Examples teach better than adjectives.
Don't stop at "I want this to sound warm and direct." Show the AI a paragraph you kept, a paragraph you cut, and the reason for each. Tell it which sentence earned your trust and which one sounded impressive without saying anything.
Then, every time you make a meaningful correction, ask yourself one question: is this a one time edit, or did I just reveal a rule about how I think? If it is a rule, save it.
That is how a file becomes a living record of judgment, and something more durable than memory inside any single platform. Models change. Tools change. Company permissions change. Your thumbprint travels.
When I returned to ChatGPT, I realized how much of myself I had already taught another system.
Months of feedback had become invisible infrastructure. The examples, the skills files, the small corrections, all of it carried pieces of my point of view. Once I saw that, I stopped treating context as setup and started treating it as an asset.
You are building that asset every time you work with AI. The question is whether it stays trapped in a chat history or becomes something you can carry, refine, and own.
Give your AI more than a task. Give it the choices behind the task. Show it what you notice. Teach it what earns a yes from you.
Then stay in the work. Your context helps the tool recognize you. Your discernment keeps the result honest. Your sazón makes it yours.
Every mind is its own world. Give the tool a way into yours.
♻️ Share this with someone whose AI outputs look polished but never quite feel personal.
¡Hasta la próxima, un abrazo fuerte! (Until next week, a big hug!)
