Whitespace is a tax on the expert
The generous consumer layout — 56-pixel rows, wide gutters, one idea per screen — is designed for someone who visits twice a week and needs to be oriented each time. A commodity trader at Northwind opens the same blotter at 07:15 and is still in it at 18:00. They do not need orientation. They need to see more rows.
On the Northwind rebuild we brought row height from 44 to 28 pixels, moved to tabular figures so digits align in columns, and set numeric fields at 13 pixels with a taller x-height typeface rather than shrinking the body face. Visible positions went from 14 to 31 without scrolling. Nobody complained about the type size; several people mentioned that they had stopped losing their place.
Density has a floor, and it is set by the error rate, not by taste. We test with the real dataset — 400 rows of live-shaped data, not eight rows of placeholder — and watch for mis-clicks and mis-reads. When adjacent-row selection errors rise above about one in fifty, the rows are too tight and we give the pixels back.
The keyboard is the primary input
Experts do not browse; they execute a known sequence hundreds of times a day. Every action on the Northwind order ticket has a keyboard route, the tab order follows the way the form is actually filled rather than the DOM order, and a command palette covers the long tail. Median time to enter a standard order fell from 38 seconds to 11.
That gain is mostly the removal of hand travel and re-focusing, not raw typing speed. It also survives training turnover, because a palette that accepts a fuzzy match on the action name is discoverable in a way that a memorised chord is not. We reserve chords for the twenty actions that genuinely recur.
The obligation that comes with a keyboard-first design is a visible, unambiguous focus state at every step. If a user cannot tell which cell is live, keyboard entry becomes a source of expensive mistakes rather than speed.
Ration the colour
A colour spent on a decorative header is a colour that is no longer available to mean limit breach. In dense operational software we hold to three semantic colours — attention, breach, resolved — and render everything else in a neutral ramp. The interface looks plain in a screenshot and reads instantly at a glance from two metres away, which is how a desk head actually uses it.
Colour never carries meaning alone. Roughly eight per cent of men have some form of colour vision deficiency, and a desk of twelve will typically include one. Every state also carries a glyph, a text label or a position, and we check the result against WCAG contrast for non-text indicators rather than eyeballing it.
Flashing and animation are reserved almost entirely for state changes the user must not miss. A row that pulses once on a fill is useful. A dashboard where six things move continuously trains people to ignore movement, which defeats the one job it had.
Show the state of the system, not a spinner
Expert users need to distinguish three things a spinner conflates: the value is zero, the value is unknown, and the feed is stale. On the Northwind price panel each source carries a last-tick timestamp that turns amber at 5 seconds and red at 15, and a degraded feed keeps showing its last value with the age attached rather than blanking. A blank field on a trading screen is read as zero, and that has cost people real money.
The same principle applies to writes. Optimistic updates are appropriate for a note or a tag; they are not appropriate for an order or a payment, where the user needs to know the server has accepted it. We show a pending state with an identifier and a time, and we make failure loud, specific and adjacent to the thing that failed.
Design against the observed workflow
Two days sitting behind users produces more design direction than a month of workshops. At Corvus Logistics the dispatch team had four windows open and alt-tabbed between them roughly 200 times a shift, because the vehicle status they needed was in one system and the customer commitment was in another. The redesign put both on one screen. That was the entire insight, and it was worth more than the visual work that followed.
Watch for the exports. Every time a user pulls data into Excel, they are telling you about a capability the interface does not have — usually a comparison, a pivot or a running total. At Corvus, three of the five recurring exports became views, and the other two were genuinely better served by Excel, which is also a finding worth writing down.