The default. Watch somebody attempt the real task, count where they stop. Remote over Zoom on the McDonald’s study, in person on ioMoVo and Toronto Cupcakes.

Fast, cheap, and best used after a live session rather than before one. An audit tells you what violates a principle, not what stops a person.

For reasoning at a scale interviews cannot reach. Useful for how often and how many. Unreliable for why, without a session behind it.

Turns a recommendation into a measurement. Only worth setting up once you have named the threshold that counts as success.

Google Analytics, to check whether a qualitative finding happens often enough to prioritise. Frequency, not cause.

Recordings reviewed afterward so the same observation is counted the same way for every participant. It is what turns eight anecdotes into seven of eight.

Writing the task so it does not contain its own answer. The hardest part of the whole method and the least discussed.

Naming things the way people ask for them instead of the way the database stores them. Three of the six case studies turned out to be a naming problem that had been filed as a navigation problem.

Clickable enough to test, not so polished that feedback becomes about the visuals. Mobile and web flows kept consistent across both.

Type scale, spacing scale, one accent, states defined once. This site runs on one twelve-column grid, three measures, three breakpoints, and an eight-step spacing scale; there is exactly one grid declaration in the stylesheet and no page sets a width in a style attribute.

Contrast checked against WCAG AA, full keyboard operability, visible focus, reduced motion respected. A requirement, not a pass at the end.

Plain language in the interface, especially in errors. Most confusing UI is a labelling problem wearing a layout costume.

Semantic markup, modern CSS, vanilla JavaScript. This site carries a little over six kilobytes of script, and no framework runs in the browser.

Deployed on Netlify, with edge functions and serverless endpoints where a page needs something the browser cannot do alone.

When an interaction depends on timing or momentum, a coded prototype answers the question and a Figma frame guesses at it.

Specifications in the units engineering uses, with states and edge cases written down. Knowing what is expensive to build changes which design I argue for.

Documentary, automotive, and brand work. Handheld and locked-off, available light first.

Portrait, editorial, event, and product. Short sessions, because people stop being themselves after about forty minutes.

Grade and cut in house. One revision round inside the agreed turnaround, delivered at the sizes the destination needs.

Getting somebody comfortable enough to be themselves on camera is the same skill as getting a participant to think out loud in a session.