Surface Tension

What it is

First you guess: in about a minute of trying to focus, how many times will you come up for air? Then you dive. A warm point of light — you — sinks into calm, deep water. The longer you stay down, the deeper you get, and only depth counts as real work.

Glints keep appearing at the bright, busy surface. Tapping one (or pressing space) pulls you up — and going back down means clawing through a murky band where the climb is slow, never quite reaching the depth you had. Meanwhile a 'busy' meter sits at 100% the whole time while the 'deep' meter barely fills.

The reveal counts it back to you: you were busy the entire minute, only a few percent of it was deep, and most of the times you came up, no one pulled you — you did. A curve drawn from the same engine shows the cliff: switch slowly and you stay deep; past a certain pace the deep work just stops. Then the reframe, and a line you can share. It collects nothing about you.

The design world — A free-dive where the warm light lives in the deep

The world is a deep, clear sea rendered in ocean blue that never becomes black. The surface is bright, choppy, scattered with glittering lures; the deep is calm. The signature inversion: the warm light is at the bottom, not the top — reaching focus warms the scene, and being yanked to the surface is the cold, twitchy place. The second inversion is motion: the descent is slow and patient, surfacing is fast and violent, and that asymmetry is the lesson — focus is expensive to build and cheap to lose.

The visualization leads and the copy is its caption: the diver's vertical position IS the focus meter, the murky band IS attention residue and re-immersion, the capped regain IS the cost that compounds. Type is an editorial Instrument Serif over a humanist Figtree — no monospace, no fine-press paper, no museum framing, deliberately none of the looks the sibling builds wear.

Unlike a focus tracker that shows you the what and leaves you feeling worse, this lets you feel the mechanism in your own hands and then hands you dignity instead of a scolding — it relieves, it doesn't try to fix you. And unlike a productivity explainer that asserts 'multitasking is bad,' the collapse here is one you cause and watch, not one you're told.

How it was evaluated

It was built eval-first: seven human-point-of-view criteria written before any code — reaches the first 'oh, I do that' on a phone in seconds; lands as relief and recognition, not shame; reads as its own thing, not a relative of any prior build; sounds like a real person wrote it; and keeps an explicit no-tracking promise.

Research-first paid off twice. Gate-0 discovery (a four-stream fan-out over the web and a knowledge base) invalidated the seed hunch — the cost of fragmented attention isn't invisible, it's unfelt and self-inflicted — and selected the dignity reframe as the real, underserved problem. Then a figure-verification pass caught two myths before they shipped: the famous '23 minutes 15 seconds to refocus' (not in any peer-reviewed paper) and '80% of interruptions are self-started' (really about half). The artifact uses only what survived.

Four adversarial reviewers, each handed the eval rubric, ran before deploy. They confirmed it reads as its own thing and lands as relief, and caught the rest: a broken keyboard path to the core interaction, a contrast failure on the key reveal number, and — most important — a Vercel analytics beacon that quietly contradicted the page's own 'sends nothing about you' promise. Removing it made the promise literally true.

unit tests on the dive engine

16 / 16

end-to-end tests (incl. keyboard-only, no-tracking)

10 / 10

research myths caught and dropped before publish

2

human-POV evals, written before any code

7

third-party requests — the gift tracks nothing

0

The whole argument is one pure, deterministic model: depth is focus, the ceiling is attention residue, the murk is the re-immersion cost. The same function draws the live dive and the systemic curve, so the chart can't lie about what you just felt.

tick(dt: number) {
  // descend — slowed while clawing back through the murk (re-immersion)
  const rate = this.cfg.descentRate * (this.inMurk ? this.cfg.murkFactor : 1)
  this.depth = Math.min(this.ceiling, this.depth + rate * dt)
  // residue: each surfacing lowered the ceiling; it recovers slowly while you stay down
  this.ceiling = Math.min(1, this.ceiling + this.cfg.residueRecovery * dt)
  // only sustained depth is real work
  if (this.depth >= this.cfg.deepThreshold) this.deepTime += dt
}

What the AI learned

• Research-first corrected the premise, not just the solution. The seed ('the cost is invisible, apps stole your focus') was half-wrong; the evidence said unfelt and self-inflicted. Building against the corrected problem is why the reframe lands as dignity instead of blame.
• Verify the famous numbers before they ship. Two of the most-repeated attention statistics are myths; a single verification pass kept them off a page that carries a real person's name.
• An artifact's own integrity eval can override a standing gate. The shelf's playbook says ship analytics; this build's whole premise is collecting nothing, so it ships none. The promise beat the metric.
• The keyboard path is part of the wow, not a checkbox. The core insight — 'you came up on your own' — is unreachable for a keyboard or screen-reader user unless coming up is a real, focusable control. A mouse-only lure would have hidden the heart of the piece from the people most likely to share it.
• Uniqueness concentrates in the hero and leaks at the edges. The bespoke dive was unmistakably its own thing; the supporting beats drifted toward dashboard convention until they were redrawn as editorial pull-stats. Counter the nearest neighbour on every screen, not just the signature one.