Everyone has the same genius

Frontier models are very strong, but they have one trait: they are universal. They are trained to be useful to everyone at once, so what they give me is what they give the colleague at the next desk. For tasks that are the same for everyone, that is wonderful. My work is different: everything I do competes with something. I look at competitors and try to do it a little better, a little differently, a little faster, a little safer, a little cheaper. Every such “little” is a decision that is not in the average answer, and it comes from experience. Everyone has their own, and so do I.

The obvious move is to wait for the next model. It will be smarter, it will arrive on schedule, and it will be just as universal. Everyone who asks it the same question gets the same smooth answer, and a product built from identical answers drifts toward every other product. A good prompt does not cure this: every morning it starts from a blank page, and the model never learns which of my past decisions it should have inherited. The agent’s built-in memory is not the way out either: the agent writes into it by itself, it stays on one machine, and after the computer is reinstalled nothing of it is left. I can neither manage it nor move it. I need a memory that I set up myself and can take with me.

Anyone can hire the genius, so only what the genius lacks will set your product apart from someone else’s: personal experience.

That experience has to live outside my head, in a form an agent can read, and still stay mine. I do not need an assistant that replaces me at work: one that writes code instead of me or does anything else instead of me. I need an exoskeleton. It makes me faster and better informed and takes the clutter off my head: the set of decisions, the reasons they were made, and the reasons the others were rejected. Freed from that, I can work in different fields, far beyond my own profession. I write this article the same way: the experience and the decisions are mine, and the agent translates, watches the grammar and counts the words, so that nobody has to read a post for half an hour. There will be many more examples ahead, but for now there are two, and both turned up by themselves while I was preparing this article, my first publication. They may not be the most telling, but they were right at hand.

The first example is the title. When I started writing this article, Claude first suggested a title and a hook, shown in the first screenshot. Would I open an article with that title and that hook? Only at gunpoint. The model does not know me, so it produced the label of a topic: a question in the title, a number in the hook, and a topic instead of a thought. I did a research pass with the scout (I will tell you who that is later) and fixed the result in skills: the stylistic form of a title and a hook, and the accepted and rejected decisions with their reasons. Thanks to this, my skill Bard (the name is no accident: every hero should have a bard of their own) will be able to write titles and hooks in the style I have adopted. I also plan to collect stand-up comedians’ transcripts, because I want bolder titles. You read the result before the first paragraph.

The second example is the site itself. I asked the model to build a page for posts and got what is in the first screenshot. I disliked it at once but could not explain why: I have never built a website and do not know the nuances. What I could do was collect experience. Two days went not into layout but into writing down what I like and why: four references with notes, one file with colors and fonts, twenty-one questions from the agent, two researchers and seven disputed points that I decided myself. The second screenshot shows the new site, whose background is the graph of my own vault. There were many decisions in this site, and I do not intend to carry them in my head: everything that will be useful next time is recorded in skills, process skills and the knowledge graph. So I can switch entirely to another field and other tasks without fearing that I will forget how a site is made. If in three months or in a year I need a site again, this one or a similar one, I will sit down and make it, building on what I have built now, and no meaning and no decision will be lost. The foundation does not crumble; with every next time it only gets better.

I am no longer at the beginning of the road, but the middle is still far off, not to mention a finished product. A noticeable part of the system already works, and I named it DIML (Drosophila-Inspired Memory Loop). It is built like this:

The vault. Notes in Obsidian, 289 of them now. But memory is not an archive where everything lies on its shelf: it has to be able to fetch what is needed and put things in order.
Three flows, like the fly’s. The signal goes into the brain, out of the brain and inside it. Afferent writes new things, but only after my “yes”. Efferent fetches what a task needs by plain search over the descriptions of the cards, with no embeddings, and runs it through the Identity section with my standing preferences, for example “do not introduce a new mechanism until the need is proven”. Intrinsic compresses and freezes matured cards at night, and that is the only thing the system does without my “yes”. On the graph these are the red nodes, and the central node Mushroom body, named after the memory center of the fly, connects them to the rest of the vault.
Skills. Some describe the order of actions, others the rules by which a result is checked. The purple nodes do not duplicate the skills themselves; they keep what the skills lack: how a decision changed, why this one was taken, what was rejected and what turned out differently than I expected.
Everything I know. The green nodes: projects, goals, resources and the Identity section itself. The agent relies on them when it solves a task.

There is one model for everyone, so a product is set apart not by it but by the decisions you made and the reasons you rejected the rest. Written down so that an agent can read them and you can dispute them, they stop depending on your memory, on your computer and on whichever model is in fashion. What happens if you arrange every area of life this way I do not know yet, but testing it is more interesting than waiting for the next model.

Everyone has the same genius, but the experience you write down turns it into your personal genius.