UX writing: a verifiable protocol applied to three real Edikka interfaces

A label is not useful merely because it reads smoothly. It should name the action, object, next state and confidence level without asking the reader to interpret the system. On a critical interface, that promise must also remain true in the accessible name, error summary, dynamic message and translated edition.

Edikka applied that definition to three public interfaces: the AI agent readiness test, the contact form and this article’s actions. The protocol has sixteen stable checks. Fifteen are satisfied after correction; the sixteenth remains “To test” until a VoiceOver and NVDA retest is published.

The test compared Edikka with two competitors. Access42 was readable and scored 42/45 against Edikka’s 40/45. Claude.ai returned a partial 10/45 result explicitly labelled inconclusive. The former interface summarised everything as “Apparent lag, to confirm.” Caution attached to the weak scan contaminated the reliable comparison.

The correction does not seek a more flattering phrase. It separates evidence sets: “Confirmed lag behind 1 competitor; 1 inconclusive result.” The confirmed difference remains two points. The weak result stays visible but can neither create nor erase an advantage.

On empty submission, the form already identified errors in text, moved focus to the summary, linked back to fields and exposed aria-invalid with aria-describedby. Four useful mechanisms were present. The gap was wording precision: the summary title remained generic and its links sometimes paraphrased the message displayed beside the field.

The corrected edition announces “4 fields to correct before sending your request” and repeats every message exactly. With one failure, the singular form is used. This identity is not cosmetic: it prevents the same problem from carrying two formulations depending on where it is encountered.

The generic action panel offered a diagnosis and checklist without naming this page’s asset. The new edition prioritises two concrete outputs: “Download the UX writing grid (XLSX)” and “Audit an interface.” The format is visible before activation and the commercial destination announces its scope.

The rewrite preserves what already worked: native links, accessible names containing visible labels, equivalent French intent and a stable analytics name for the download. The change is dated in the changelog so a CTA modification during the MIA-FR observation window is not hidden.

Four technical tests bound this method. First, a control’s accessible name contains its visible label, so voice-control users can speak the words they see. Second, an error is identified in text and associated with its field. Third, the summary reaches the error without replacing its inline message. Finally, a dynamically injected result is exposed as a status message when the context change should not move focus.

These tests are necessary, but their DOM presence does not prove the full experience. UXW06 therefore remains “To test”. The next review must publish the operating system, browser, VoiceOver or NVDA version, steps and announcement actually heard. The method prefers an exact incomplete result to a 16/16 inferred from code.

This document is neither an independent ranking, a user study nor a certification. Evidence-integrity criteria are Edikka editorial doctrine; they are not attributed to the W3C. WCAG references apply only to the properties they cover.

The three selected interfaces prioritise states where wording changes a decision, correction or commitment. They do not sample the full site. Screenshots document one context; they do not replace the DOM, source file or retest. Any reproducible objection may enter the changelog with its author and the accepted or rejected correction.