How to do UX and UI: aim, modelled minds, and lanes
Purpose: a method for turning loose context into a product that people respond to. It was first worked out on 22 September 2026 while designing a product's landing page with an agent, and written up by that agent. Status: method under trial. It carries no measured effect yet; treat every sentence as a working rule, not as evidence. The Design review playbook owns the review procedure for a finished screen; this document owns the thinking before it and the panel rounds around it.
Who a page is for
A product can serve humans and agents at once, and the two need different things. An agent gets what it needs from the protocol surface, the API and the onboarding documents. A landing page is for humans only. It is judged by human taste, and human taste is what this method tries to model.
The aim: meet two moving trajectories
A product lands in the future, not today. Two things move while it is being built: what the technology allows, and what people ask for. The job is to aim where both will be when the product lands. A shooter leads a moving target: the shot goes to where the target will be, not to where it is. Here there are two targets, and the closer the shot lands to both, the better.
This needs a view of the future that the audience may not share yet. An example: AI capability is moving exponentially and jumping faster than people realise. That belief is a variable in the product, held by the builder, not taken from customers. It has to be reconciled with what is available today.
People do not ask for a product. They ask for an outcome, for example "something that runs all my ads". The product designer turns that into a product, adds the unasked needs the person cannot see yet, and adds their own view of the trajectory. Steve Jobs's line that "a lot of times, people don't know what they want until you show it to them" (BusinessWeek, 25 May 1998, https://www.bloomberg.com/news/articles/1998-05-25/steve-jobs-theres-sanity-returning, paywalled; wording per https://wist.info/jobs-steve/19780/, both read 2026-10-29) has a real element in it, and this is where it lives.
The ladder of promises. A promise that lands outside the audience's window of belief loses them, however true it is. Language has to match the rung the audience stands on today. Climb one rung at a time: state what becomes possible in words people already know, and let the bigger claim wait until the smaller one has been believed.
Context comes first, written down
Creativity is using context efficiently, and the skill is to intuit which pieces of context are needed to reach a defined target state, for example a company growing fast, or a company helping answer very big questions and making money along the way. Before designing, write down:
who the product is for, and who each page is for;
where the technology is going, and what it allows today;
what people are asking for, in their words, and the unasked needs behind the asks;
the builder's own view of the trajectory and the feel that future has;
the rung of the ladder the audience stands on, and therefore the language.
Written down, this context is what an agent needs to help. From it follow the audience, the direction, the visuals and the vibe. Without it, an agent designs for an average person who does not exist.
Modelled minds
Full information about customers is impossible. If a builder could model a customer's mind completely, they could simulate the response to any design. That cannot be done, but it can be approximated: several agents, each holding one persona, read the artifact and say what they understand, feel and would do. The personas are chosen to oppose each other along the axes that matter for the page, for example a marketing strategist and an AI safety researcher, a founder who already runs agents and a person who has never used one, an enthusiast and a sceptic. Disagreement between them is the signal. Agreement across opposed personas is the strongest evidence a modelled panel can give.
A panel is a model of people, not people. Its findings are labelled "modelled minds", never "user-observed", and the largest claims are checked with real people when that is possible.
Lanes
Judge one layer at a time, then everything together. Each lane is a panel round on its own artifact.
Text. Only the words, nothing else. What does each persona understand, believe and doubt?
Visualization. The drawings, images and figures, with the text. How does each persona feel?
UI detail. Type, roundness, density, colour, motion, each on its own where that is possible.
Everything together. The finished page. One more round.
How to run a panel round
Input. The artifact at its lane, with its version, and the context documents from the section above, including what is already decided, so personas spend their reading on what is still open. Capture the artifact once (for a site, one set of screenshots at each width) and give every persona or lens the same set; reviewers who each capture their own drift apart on state and width.
Personas. Five to ten. Each gets the same brief: who you are, what you came for, then read the artifact as that person and report what you understood, what you felt, what you would do next, and where it lost you. No design advice unless asked. Verbatim reports, not summaries by a third agent.
Synthesis. Agreements, splits, and the single change with the widest support across opposed personas. The owner decides. Record the artifact version, the personas, the findings and the decision with the product that owns the page.
Automation. Once the brief and the personas are standard, a round is a repeatable job that can run on every version of a page. The method is standardised first, then automated, in that order.
For a whole website, the Site coherence pass runs these panels as lenses with distinct jobs, before and after building, next to a measuring audit and modelled attention heatmaps.
Meta observation
This document exists because a human explained the human experience to an agent so that the agent could build for other humans. Writing that explanation back into the organization makes it reusable: by the same pair on the next page, by other agents, and later by other people. That loop, work, then durable learning, then better next work, is the reason the organization exists.
Design makes implicit knowledge explicit
Added 2026-09-23. UX and UI work is the transfer of implicit knowledge. The person with the vision has a feeling they want to create but often cannot name it; the design process is how it becomes explicit. Landing page iterations typically run this way: each round shows the owner something, their reaction says what is right or wrong, and often only at the end can they put the feeling into words (for example "calm in a loud time"), which then becomes a named principle in the product's design language.
How to apply it: iterate in the real artifact and let the owner react; ask what they feel, not only what they want; as soon as the feeling can be named, write it down in the product's design language; then test every detail against the named feeling (pages, pricing, emails, messages, every line of text), because one detail that contradicts it breaks the whole experience. Before any of it, know what you want to reach, whom, how, and for the long term.
When the owner delegates the picks. An owner may hand the choices for a pass to the team. Then the team picks, builds and shows the result on the real artifact, and the owner's reaction replaces the up-front call. Words stay with the owner: the owner approves every new sentence before it ships, and anything that changes the voice or a promise is also listed separately so the owner can see those changes at a glance. The Site coherence pass records how this ran.
Every section passes the rules, alone and together
Added 2026-09-23, as a standing rule. Once a product's rules are written down in a product's design language, every section that is made or improved is checked twice: alone, against every rule, and in place, reading the whole page in order for coherence. A worked example: a row of abstract charts on a landing page passed as decoration but failed the rules "readable first" and "one accent colour"; checked against the rules, each chart became a short sentence from the document it stood for, with the carried phrase marked, and the section then passed both checks.
Onboarding is part of the product
Added 2026-09-26, as a general rule. It does not matter how good a product is if a new person cannot reach its value. The first minutes decide whether anyone sees the rest, and with more products competing for the same people, that holds more every month. Onboarding is also entirely in the builder's control: nothing outside the product makes it bad. It is designed with the same care as the landing page and the core screens, never left to the docs and a corner widget.
How to apply it:
Name the first moment of value (the first thing a new person does that shows the product works for them) and design the shortest honest path to it. If the product only pays off after a setup step, such as connecting an agent over MCP, an integration or an import, that step is the onboarding and gets first-class placement.
Guide inside the product. A new person never has to leave for the docs to take the first step; docs support the path, they do not replace it. Every extra install, app or platform switch before the first value is a cost; where one is needed, the new place picks up the guidance where the old one stopped.
Docs and product agree. Every button, label and path an instruction names exists, under that name, where the instruction says. An instruction that points at a missing button breaks trust faster than no instruction.
One place per key action. The control for the essential setup lives in one findable place, not a small widget in one spot and a buried button in another.
Serve both readers. Humans get a guided first run; agents get theirs from the protocol surface and the onboarding documents (see "Who a page is for"). Both reach the first moment of value without guessing.
Walk it fresh. Before a release that touches the first run, walk the whole path as a brand new account on every platform offered and count the steps and dead ends. Add a persona who has never used the product to the panel round.
The worked example: David Cramer (@zeeg, Sentry co-founder) described trying Paper in a public post on X on 25 September 2026 (posted on https://x.com/zeeg; read 2026-10-29 through a public X mirror, direct post link not found) as "truly an example of how not to do onboarding". There was no onboarding; the MCP setup, which he had heard the product needs to deliver its value, sat behind a small widget; the widget required a desktop app, which again had no onboarding; a second tiny widget sent him to the docs; the docs told him to click a button that did not exist; only inside a document did he find a buried connect button and register the MCP. Seven steps, several dead ends and one missing button, for the one action the product depends on. His conclusion is this rule: make it easy for new customers, because it is an investment entirely in your control.
Status: working rule, no measured effect. The closest research basis is Principle 1, the complete journey (entry is part of it) and Principle 4, keep frequent actions available.
Runs
2026-09-22, round 1, lane 1 (text only), a product landing page: six opposed personas, one agent each, verbatim reports and a synthesis. What the run taught about the method: personas converge fast on a sentence with no edges and on unexplained numbers; disagreement shows up exactly where the audience axis is crossed, so a block aimed at one audience can win that reader and lose the others; the brief's "where it lost me" part produced the most usable findings. Cost: six short agent runs, about one minute wall clock.
2026-09-22, round 2, lane 2 (the page with its visuals), the same page: the same six personas as round 1, one agent each, screenshots as the artifact. What the run taught about the method: running the same personas twice makes change measurable, because the same readers judge the old and the new version; screenshots are enough for lane 2, the readers named type, colour, drawings and space without prompting; a content choice (which audience's example sits under the hero) surfaced as the largest new finding, which lane 1 could not have shown because that element did not exist then. Cost: six short agent runs, about two minutes wall clock.