almost
yours

Personalization, as software has mostly practiced it, means the system comes to know something about a person. It watches what they click, what they pause on, what they buy and skip and return to, and then it quietly rearranges the world in response. This is the familiar logic of implicit feedback and recommender systems. The feed becomes more accurate. The recommendation improves. A default changes somewhere, and a surface begins to anticipate.

There is real value in this. I do not want to pretend otherwise. But it is a narrow account of what personal can mean, because being predicted is not the same as being expressed. A recommendation can reflect a preference without helping me understand why I hold it, how I would revise it, or what it says about how I see. The system knows me from the outside.

The question I keep returning to is different. Not how a system can know a person better, but what kinds of structures make a person want to leave something of themselves behind. Personalization not as prediction, but as participation.

It becomes more interesting when it stops choosing on our behalf and starts giving us structures we can make our own.

A template, a prompt, a crop, a checklist, a grid, a preset, a style. We usually file these under features. I have come to think of them as invitations. They give a person a way to begin, and beginning, it turns out, is most of the problem. In HCI, William Gaver's work on affordances is one way to say this more formally: the designed thing does not merely sit there. It suggests possible action.

We tend to talk about creativity as if it starts with openness. A blank page, a blank canvas, infinite possibility. But infinite possibility can feel strangely indifferent. A blank page does not always feel like freedom. Sometimes it feels like the absence of help.

Most people are more creative than software assumes, and less willing to begin from nothing. I do not think the problem is that people lack taste, judgment, or imagination. It is that these faculties need something to react to. Taste is mostly a responsive instrument. It says almost, not that, closer. But it needs a first move to answer.

Structure changes the emotional state of creation. It lowers the cost of the first move so the person can make the second. This is why templates work, and presets, and prompts, and recipe formats, and moodboards, and character creators. They are not fully open, and that is exactly why they are useful. Don Norman's account of constraints in design points to the same thing: limits can make action intelligible. Good structures offer enough constraint to reduce the fear of beginning, and enough room for the person to appear.

None of this arrived with AI. Good tools have carried this quality all along. What has changed is that the quality is moving from the background of software toward the foreground.

Some years ago I spent a lot of time in ultrasound labs, watching sonographers work. What struck me was how rarely a sonographer knew the image in advance. They found it by adjusting: gain, depth, focus, a slow sweep of the transducer, and then a reaction to what came back. Too dark. Too noisy. Not the plane I need. Closer.

The control panel was not a set of settings to be configured correctly. It was a structure for negotiating with the image until judgment could get a grip on it. Decades of accumulated practice were encoded in those controls, and every scan was a small conversation with that inheritance. This is close to what distributed cognition has long tried to describe: thinking is not sealed inside the head; it is carried across people, artifacts, representations, and practiced routines. The machine made the first move, over and over, and the sonographer's expertise lived in the corrections.

Note-taking tools work on the same principle, though we rarely say so. A blank note is freedom, but also a kind of emptiness. A folder gives a thought a place; a tag gives it a relation; a checklist gives it sequence; a backlink suggests that a thought might not be alone. The person arrives with fragments, and the tool offers structures of continuation.

This is why note-taking apps are never really about storage. They are about the hope that thinking has an architecture, and that if we choose the right structure, something in the mind will become more visible to itself. Work on personal informatics makes a neighboring point: systems that help people collect, integrate, reflect, and act on traces of their lives become part of how those lives are understood. Whether that hope is ever fully repaid is another question, but the reaching is real.

I felt this most directly while building a classification system for customer feedback. The taxonomy I started with was reasonable. It would have looked fine in a review. It only became mine through the inputs that refused to fit.

Every misfit forced a decision about what a category actually was, what counted as one thing rather than two, what deserved to be tracked at all. The final structure carried those decisions the way a well-worn tool carries the grip of its owner's hand. Nobody could have extracted that schema from my behavior. It had to be produced, argument by argument, against material that pushed back.

Even the spreadsheet, which we treat as the least expressive software imaginable, has this quality. The grid asks the world to become rows and columns. It turns ambiguity into comparison. It forces decisions about what counts as a category, what counts as a value, and what can be ignored. Two people modeling the same problem will produce two different sheets, and the difference between them is not noise. It is a record of how each of them sees.

This was personalization before we called it that. Not the system adapting invisibly to the person, but the person adapting a structure until it begins to carry their way of seeing.

There is a specific moment in good tools that I find myself watching for. It is the moment when the tool gives something back that is close, but not quite right.

The generated paragraph that almost sounds like you. The edit that almost has the right mood. The template that almost fits what you are trying to say. That gap is not only a failure of the tool. It is where the person enters.

When a tool returns something imperfect but legible, it creates a surface for correction, and correction is where preference becomes visible, often for the first time even to the person doing the correcting. Karl Weick's work on sensemaking is useful here: people often understand what they think through what they find themselves saying and revising. The personal is not always waiting inside us to be extracted by better data. Sometimes it is produced in the interaction between a person and a structure.

You do not know what kind of room you want until you start moving furniture. You do not know your writing voice until something gives back words that are almost yours, and the wrongness of them tells you something the rightness never could.

Michael Polanyi's idea of tacit knowledge helps name the other side of this. We know more than we can say. What I would add is that structures are one way tacit knowledge becomes actionable. The sonographer may not be able to write down every property of a good image, but give her a gain knob and she will find it. I cannot fully articulate my own writing voice, but show me a paragraph that misses it and I can tell you where, and slowly the wheres accumulate into something like a description.

This is where AI changes the texture of the problem. If AI is treated only as an output engine, personalization becomes a better autocomplete: faster content, more variants, more things that resemble what the system guesses we might have wanted. Useful, but still the old idea wearing new capability.

The more interesting possibility is AI as a structuring medium. In older tools, the structure was fixed and the person personalized inside it. The crop tool did not learn what you tended to cut away. The folder system did not notice which notes kept resisting their folders. In AI tools, the structure itself can respond. A prompt can become a template. A template can become a conversation. A conversation, over enough rounds, can become something like a personal grammar.

I have watched this happen in my own writing practice. What began as instructions to a model slowly became a governing document: a small file of rulings, accumulated one rejection at a time. This construction is banned. This move has a name now. This register, not that one. None of it existed before the tool gave me things that were almost right. Every ruling was extracted from a specific wrongness.

The document is now among the most personal artifacts I own, and I did not write it so much as arrive at it. The structure and I changed each other.

There is a caution worth sitting with here. A structure that responds too eagerly can close the gap entirely. The paragraph arrives already right. The taxonomy arrives already finished. There is nothing left to push against. If the gap is where the person appears, a system that eliminates every gap may be eliminating the person along with the friction.

The design problem is not to make the first move perfect. It is to make the first move productive: close enough to react to, imperfect enough to require a reaction.

The distinction I find useful is between relevance and authorship. Relevance says: this matches you. Authorship says: this contains you.

Most personalization systems have been optimized for relevance, and they are genuinely good at it: selecting, ranking, recommending, predicting. But authorship is the stronger form, because it involves the person leaving a trace on the system. A personalized feed may know me, but I did not make it. A marked-up document, a tuned prompt, a taxonomy argued into place against unruly material, a note system bent around one mind's habits: these contain decisions. They become personal because someone acted on them.

Less like a mirror, more like an instrument. A mirror shows you what you already are. An instrument gives you a structure through which you can become audible. Nobody would say a cello knows the cellist, though after enough years it fits no other hands quite the same way.

I do not know exactly what this next generation of tools should look like. But I suspect the products that come to feel most personal will not be the ones that know us so completely that we no longer need to act. They will be the ones that give us enough structure that we instinctively begin to make something our own: a first move we can push against, a gap we can enter, a better loop to step into.

A product becomes personal, in the end, not when it removes the person from the loop, but when it gives them a better loop to enter.