Quiet Bands designs and builds custom software, internal tools, integrations and automation around operational problems, from Houston. The method is to understand the work as it actually runs before deciding whether software belongs in it. Sometimes the finding is that nothing should be built.

Built 001 · in production

Frenchies Coffee Bar, Connecticut. Order ahead, rewards and payment for a two-location coffee company, built to run at counter speed without adding work behind the bar or routing regulars through a marketplace built for someone else's economics.

One shipped product is listed because one has shipped. The count will change when the count changes.

Building software got easier. Knowing what deserves to be built didn’t. So every idea is tested before any of it is built, and the test is written to be able to say no. The page you came from does it to a lightbulb.

Six things we look for, in order

1. Problem. What goes wrong, for whom, and how often. A problem that happens rarely rarely deserves software.
2. Current behavior. What people do today instead. The workaround is the strongest evidence that a problem is real.
3. Consequence. What it costs: time, labor, revenue, errors, risk. No consequence, no reason to change.
4. Evidence. What you have actually seen, as opposed to assumed.
5. Human validation. Whether the people doing the work have told you any of this themselves.
6. Decision. One of five: research more, talk to humans, test manually, explore a build, or park it.

Most ideas that come through here should not be built yet. Quiet Bands builds software for a living, and the review is written so it can still say so.

Take the Pressure Test Read the Field Guide (PDF)

A loading indicator, October 2026. Live. Hold one at /experiments/thinking-orb/.

A wait that shows what kind of work is happening: a globe scanned for searching, a constellation wiring itself for connecting, bands that scramble and click back for solving. Nine states, taken off React onto plain canvas, with one thing the original doesn't do: you can pick it up. Hold it and it stops, drag it and it turns through its own motion, let go and it coasts back to its own pace. It already runs on the Work intake button.

Hold one Watch the reel (15 s)

A homepage experiment, September 2026. Retired. The code stays at /lab/spatial-hero/.

The thesis

The screen is a window into a volume. Nothing is painted on the glass; every object has a depth, and you perceive the depths through differential motion, occlusion and scale. The line Request — Complete sits at depth zero and never moves. Six words stand edge-on in it, invisible from the front. Tilt the phone and they turn open and fan out by depth. Return to the front and they close to nothing. The claim was that the hidden steps of a process could be shown, rather than listed.

The diagonal problem

A phone in the hand always carries some pitch. Each deeper word moved further down the screen for it, so the six words drew a staircase down a diagonal instead of a row standing in the line. The vertical parallax on the words was removed; they kept one Y origin and separated only by X, depth, scale and occlusion. The same pitch also pushed the wordmark into the rule. That was fixed by lowering the pitch response, not by moving anything, because no placement change fixed it without breaking the composition at rest.

The stratum test

Version 1.3 packed the six words onto the rule itself, so that at an angle the hidden language separated along the Request-to-Complete axis, nearer words occluding deeper ones. At full tilt: a band of activity on the line. At six degrees: a row of slivers. At rest: the rule. It worked as a mechanism. Whether it read as a process or as scatter was the open question, and the honest answer was that the words were doing the work.

The geometry test that failed

So version 1.4 removed the words and left the surfaces: six empty outlined planes, edge-on in the rule, opening with the angle. If geometry alone could carry "concealed structure", this would show it. It did not. At six degrees the row of small parallelograms read as tabs or pagination. At full tilt the overlapping outlines read as wireframe panels, which is to say cards. In the scroll passage the receding stack read as depth, but as a deck, not as a process. A drawn box is a card. Geometry in that form carried "six things at six depths" and nothing more.

Why the motion worked

The input side held up. Tilt, thumb and scroll all wrote one viewing angle; a critically damped spring carried it, so a release kept the finger's velocity and an interruption re-targeted from the live value; past the edge it rubber-banded. Scroll was a dolly, and the camera drifted slightly aside on the way in, so scrolling alone opened the words before any tilt did. The only real bug in the motion was a permission prompt that gave up after one attempt on iOS; a scroll's touch end carries no user activation, and the request had to keep asking until Safari answered. That engine was kept and is what drives the entry you came from.

Why it was retired

Because the meaning lived in the words, and the words could not survive the test that mattered: take them away and the object stopped saying anything. A homepage that depends on six specific nouns is a slogan with a parallax effect. The mechanism was sound; the object was not. It was replaced by an object that says the same thing with no words at all, a bulb that breaks into the site.

Measurements quoted in the lab notes were made in headless software rendering, not on a phone. No user testing was done. Treat every judgement above as one person's reading of the frames.