Beyond Slop: Critique, Audit, Hardening, and Adaptation
A clean detector run closes only a narrow question: did the configured mechanical rules find a primary problem in the evidence they could inspect? It does not establish that a surface is specific to its product, that a first-time user understands the next action, or that the interface survives real data and devices. Impeccable's other quality workflows exist because those are different failure surfaces. The practical task is to diagnose the problem first, then choose the pass that can actually answer it.
Critique asks what the rules cannot
critique separates design judgment from deterministic evidence. When subagents are available, its method calls for two independent assessments. One examines design specificity, hierarchy, information architecture, cognitive load, emotional journey, personas, and Nielsen's heuristics. The other runs the detector and browser evidence. The results are synthesized rather than concatenated, so the report can say where human review and mechanical evidence agree, where the detector found something the reviewer missed, and where a rule is a false positive.
That separation matters because a page can be mechanically clean and still be wrong for its audience. Imagine a landing page with strong contrast, no overflow, valid headings, and no obvious generated-UI tells. If its first action is buried, its copy assumes knowledge the visitor does not have, and its composition could belong to any competitor, a detector has little to say. Critique can examine those failures through design specificity, cognitive load, the emotional journey, and relevant personas.
The personas are not decorative character sketches. They are ways to test different failure conditions. A power user asks whether the main task is efficient and whether shortcuts or batch actions exist. A first-timer asks whether the first action is clear and whether jargon blocks progress. An accessibility-dependent user exposes keyboard, focus, semantic, contrast, and state-announcement failures. A stress tester pushes long strings, empty states, unexpected inputs, refreshes, and concurrent actions. A distracted mobile user tests interruption, thumb reach, slow loading, and touch targets. The relevant persona depends on the surface; a landing page and a data-heavy dashboard should not be tested through the same expectations.
Critique also distinguishes the complexity inherent in a task from the complexity created by the interface. A complicated design-review workflow may carry unavoidable intrinsic load. Unclear labels, inconsistent patterns, unnecessary choices, and context switches add extraneous load that the product can remove. Progressive disclosure, grouping, visible context, and feedback can then support the useful effort of learning the product. This is why cognitive-load review belongs beside visual hierarchy rather than after it: the question is not only what the page looks like, but what the user must hold in mind to act.
Heuristic scores and P0 to P3 priorities structure the report, but they are not scientific measurements of product quality. A score can make a weakness visible and comparable within one review. It cannot turn an inherently contextual judgment into an objective fact.