Principles
We are living the AI-First moment. Interfaces are being rebuilt around what the machine can generate, predict, and automate. Something important is being decided, mostly without discussion: what kind of space a person occupies when software mediates their attention, their choices, their relationships.
The pain, today
None of this is five years away
The long argument about what this technology does to attention and agency will take years to settle, the way the argument about social platforms took years. This is not that. These are the things it already costs, now, to the people furthest along in using it.
-
Checking costs what doing costs
The output cannot be used unsupervised, so all of it gets verified. Verification is the work you were trying to skip, and it arrives at the same volume you were trying to avoid.
-
The agents got faster and you got busier
Throughput went up. So did the part that was yours: reloading context, composing the prompt, deciding what was actually true. The bottleneck did not disappear. It moved into the person.
-
Every session starts from nothing
The same context gets re-explained every morning. The system keeps none of the relationship, which means you keep all of it, and the tenth conversation costs exactly what the first one did.
-
You accepted the suggestion, and something shifted
Hard to name and easy to feel. The work went out under your name, and you are no longer certain which parts of it you decided.
-
Confidence arrives without warrant
The answer is fluent, well formed and complete. Nothing in its shape tells you whether it is grounded, so you find out yourself, every time, at your own cost.
-
The stack is relearned every quarter
Each release moves the surface, rebuilt again rather than carried forward. Time goes into re-acquiring what you already knew, and what you learn does not accumulate into anything.
Why it happens
The person's part was never designed. It was left over.
None of those six is a defect, and none of them is anyone's bad intention. They are what a system produces when it is organised around what the model can do. The machine's territory is drawn first, as wide as capability allows, and whatever it could not reach becomes the person's job. That boundary was not chosen by anyone. It fell out of a limitation.
The same system, three model releases apart
Which means it moves. Every capability jump redraws it, always in the same direction, and the person's part does not get smaller because someone decided it should. It gets smaller because the residue shrank.
That is a stability problem before it is anything else. A system whose human boundary is residual has to be rethought at every model release, and the people who work inside it have to relearn where they stand. No roadmap has a line for that, and the cost is paid anyway.
The position
Human-first AI
Not less AI. The systems here run fleets of agents and automate hard, further than most, and the machine is given everything it can carry. What changes is the order of two decisions. The person's part is settled first, deliberately, and the machine gets all of the rest.
The line between what the machine does and what stays with the person is chosen, not residual.
Where the line falls is not fixed
There is no single correct placement, and this is not a doctrine with one answer. Halo draws it at authorship: the system can remind, it cannot decide what the protocol says. Compass draws it at decision and presence, and hands over research, tailoring, tracking and memory without hesitation. Prisma draws it at reading the divergence, and does the triangulation itself.
What those share is not where the line sits. It is that the line was a decision, and that you can be shown it.
The seven
Seven places the line has been drawn
They are a set rather than a sequence, and they are older than this argument. Each one is a position the line has been held at, and each names the work that puts it under load. Where no single work can carry the claim, it points at all of them instead.
Presence over noise
Interfaces must create space for attention, not consume it. The attention economy has made distraction a business model and interruption a design pattern, and both are an architectural failure. The measure of a good interface is not time spent in it, but the quality of presence it makes possible.
Continuum runs fleets of agents on real work and then refuses the metric that would flatter it, measuring not how much ran while you were away but how little of it you had to hold.
Clarity over confusion
Simplicity is orientation, not minimalism. A person is clear when they know where they are, what is happening, and what their choices mean. The work is wayfinding: structure that lets a person locate themselves at every moment. That structure starts below the interface, in the data itself.
PrismDB makes the case below the interface, where nobody is watching: a record has fields, and a search that flattens them for the convenience of the index has thrown away the only thing that made the record legible.
Autonomy over manipulation
Suggestion is not control. Design must restore agency: honest options, legible consequences, and the user's intention treated as the system's instruction, not its raw material.
Halo holds a protocol it is not permitted to author. The gates between phases wait for a clinician rather than for a date; the system can remind, it cannot decide.
Coherence over novelty
Systems must feel whole, even while evolving. Novelty is cheap to ship and expensive to inhabit. Evolution should feel like a building being lived in and cared for, not demolished each season for the sake of a changelog.
All seven, together are the only evidence for this one. Systems built years apart for unrelated problems keep arriving at the same three planes and the same plain files on disk, a shape nobody redesigned for a season. It is a property of the set, and no single work can show it to you.
Relational intelligence over isolated response
A question asked at midnight is not the question asked at noon; the tenth conversation is not the first. Systems that treat every exchange as stateless force the person to carry the relationship themselves. Intelligence that matters is relational.
Compass makes the recruiter permanent and the vacancy disposable. A contact outlives every campaign they ever brought, so the tenth message arrives with the previous nine already attached. And the person is not the one holding them together.
Symbolic integrity
The interface is language; its form is never neutral. The symbolic layer carries responsibility of its own: the interface should say, in its form, exactly what the system means, nothing else and nothing less.
Prisma treats a single fluent paragraph as a form making a claim about knowledge it cannot back. So the output is typed instead: convergence, tension and blind spot are structurally different objects, not three tones of the same confident voice.
Silence as affordance
A good interface knows when not to speak. Restraint is a capability, not an absence: the notification not sent, the suggestion withheld, room for the person's own mind.
Reloop holds a thought while it ripens and shows the person nothing during the wait. If the screen is quiet on a Tuesday, either nothing arrived or everything was handled. The emptiness is the report, not the failure.
What it costs
Deciding the line first forbids things that are easy to ship. The demo is slower, because the system stops and waits where a bolder one would carry on. The metric that would make this work legible to whoever funds it is the one being refused. And a capability that could be absorbed is deliberately not absorbed, which reads as a missing feature right up until the moment you need the judgement it was protecting.
The rule underneath it: a person keeps whatever they would have had to do for the result to be theirs. Not a hypothesis this page arrived at. It is the standing principle the systems above have been held to for years, one decision at a time, and the seven are what it looks like under load.
Where it is tested
Each principle names the work that puts it under load, and each of those works draws the line somewhere you can be shown. They are not illustrations of the argument. They are the argument, running, with the cost already paid.