Interface design process

The interface design process is the sequence of steps by which an interface is understood, sketched, shown to people, built, checked and revised, together with the question of which of those steps change how usable the result is. The accepted framework is human-centred design, in which the context of use and the user requirements are specified first, design solutions are produced, and the design is evaluated, and the cycle repeats (International Organization for Standardization, 2010). Beyond that framework the evidence is uneven. Showing several designs before committing to one has controlled support, though not in every task. Personas (descriptions of imagined users) and task scenarios have mixed results and no direct measure of the usability of a finished product. Specifying every state of an interface, review by someone who did not build it, and the use of AI design and coding agents rest mostly on indirect or preliminary studies. How people best express what they want visually is thinly studied. Methods for evaluating an interface once it exists, such as user tests and heuristic evaluation, are covered in Usability evaluation.

History

ISO 9241-210:2010 sets the vocabulary this article uses. It names understanding and specifying the context of use, specifying user requirements, producing design solutions and evaluating the design as the design activities, and says that the process is iterative and driven by user-centred evaluation (International Organization for Standardization, 2010).

The history of user-centred design before the 2010 standard, and of scenario-based and parallel design, is not covered by the research.

Core concepts

The cycle and its iterations

Nielsen Norman Group separates formative evaluation, the iterative improvement of a design before production, from summative evaluation, the comparison of a shipped product with a benchmark (Kendrick, 2019). The process described here is formative. Studies of how much a round of redesign gains give figures that depend on the case. Nielsen reported that measured usability rose a median 38% per iteration across four case studies, with progressively smaller gains in later iterations (Nielsen, 1993a). In a controlled comparison of four designs, one iteration gained 12 to 29 percent, averaging 18, which is less than the earlier survey (Nielsen & Faber, 1996).

Gains are not guaranteed. In the Census Bureau’s American FactFinder redesign, task accuracy in iteration 3 (52%) was lower than in iterations 2 and 2.5; the problems were so large that earlier fixes could not be evaluated, and the team said further testing was needed to see whether fixes raise performance or cause new decrements (Olmsted-Hawala et al., 2011). A two-round library website study (8 students) found that the second round caught a problem that the changes between versions had created (DeStasio & Jeitner, 2024). A newsletter summary of Tan et al. (2001) reports that a Web site retest found fixing problems created 9 new problems and cut the problem count by only 37% (Bailey, 2002).

Briefs with users and scenarios

Design briefs often describe the intended users as personas and the work they come to do as task scenarios. The persona literature reports benefits in some conditions. Long reported designs with superior usability characteristics in a 5-week student experiment (Long, 2009), and Miaskiewicz and colleagues report that personas lead to more effective designs when empathy for the persona is created (Miaskiewicz et al., 2009). Other studies find less. In a randomized experiment with 79 students, persona models lowered the priority given to antisocial user stories but had no significant effect on prosocial stories (Ayoola et al., 2024). In an exploratory study of heuristic evaluation, evaluators with personas reported more navigation findings and those without reported more design findings, and evaluators who built their own personas included more complex issues (Friess, 2015).

The persona evidence is mixed and mostly indirect. In a 60-student experiment, students primed with one or several personas generated more ideas, and the multiple-persona group more original ones, than a control, but the outcome was ideation, not usability, and perspective taking did not differ between groups, so empathy as the mechanism was not shown (Dhadphale, 2023). In an experiment on personas of varying attractiveness the solutions designers produced mostly did not differ by attractiveness (Salminen et al., 2023). In practice, interviews with four professional developers found they were not using personas as the literature describes (Billestrup et al., 2014), and a survey of 203 practitioners with 26 interviews reports that how often personas are used and how effective they are varies across projects (Wang et al., 2024). For scenarios, a with-and-without quasi-experiment on an engineering trade-off (not an interface) found that scenarios improved most decision-making practices with a small effect size (Parolin et al., 2023).

Showing alternatives before building

Parallel prototyping means producing several designs at once, getting feedback on all of them, and then refining. In a randomised web-advertisement experiment (33 participants), both conditions made five prototypes in the same time; the parallel condition made three, got feedback on all, then made two more. Parallel ads had higher click-through rates and higher expert ratings, averaging 24.4 against 21.7 out of 50 for serial ads (Dow et al., 2009). In an 84-participant ad study in pairs, sharing three designs, rather than only the best one or only one, improved outcome, exploration and critique quality (Dow et al., 2011). An independent lab found parallel prototyping gave greater design success than iterative prototyping among first-year undergraduates in a physical design competition, and the participants did not notice the benefit (Murphy et al., 2022).

Showing alternatives also changes the feedback. In a 48-participant usability test, each design scored higher on average when seen alone than when seen among three alternatives; the difference was significant for two of the three designs, and negative comments differed significantly for one (Tohidi et al., 2006). The same study found that testing with users, even with several designs, does not produce constructive suggestions for improvement: it identifies problems, not solutions (Tohidi et al., 2006). The size of the effect, and its cost, are disputed (see the controversies below).

The fidelity of what is shown

A direction can be shown as a sketch, a wireframe, a polished mockup or a working page. Controlled studies mostly find that rougher versions surface the same usability problems. Low- and high-fidelity prototypes found largely the same problems in two experiments with 10 subjects per condition; some problems appeared in only one condition, which the authors attribute mainly to small-sample variability, without direct evidence (Virzi et al., 1996). In a 48-user study, reduced-fidelity prototypes (paper, a 3D mock-up) generally predicted the usability of the real appliance, although test implications depend on fidelity, and experts reported more usability problems than novices but of lower severity (Sauer et al., 2010). A mobile-phone study dissents, finding significant differences between low- and high-fidelity prototypes for each factor (Zhang et al., 2012).

Fidelity changes other things. In a controlled experiment, task completion time may be overestimated with a computer prototype, and users overrated the aesthetic qualities of reduced-fidelity prototypes (Sauer & Sonderegger, 2009). Participants shown sketchy concept representations gave more elaborate feedback, with suggestions that referred to their past experience, than those shown refined ones (Ozcelik Buskermolen et al., 2015). Children rating one game concept in wireframe or photo-realistic form did not rate the two forms differently, though that paper notes that fidelity studies with children give contradictory results (Sim & Read, 2024). A practitioner summary reports, second-hand, that Walker et al. (2002) found no difference in the number of usability issues between paper and computer prototypes but that fidelity changed the types of issues, and that Sefelin et al. (2003) found participants criticised equally on paper and computer though 92% preferred the computer prototype (Sauro, 2017).

The fidelity evidence concerns usability tests of small products; whether it holds for AI-generated pages is not established. Part of it is reported second-hand, and the one finding on feedback type (Ozcelik Buskermolen) concerns concept representations rather than interfaces.

Walking scenarios and specifying states

An interface has many states beyond the ideal one: empty, loading, error, long content, a small screen. Indirect evidence on designing each one in advance comes from neighbouring fields. In a study with experienced engineers, stepping through scenario steps found up to three times as many exceptions as examining the user requirements (Alexander, 2000). A taxonomy of bugs in LLM-generated code includes a Missing Corner Case pattern, in which code works except for overlooking certain corner cases (Tambon et al., 2024), and a benchmark of interactive webpages reports that multimodal models generate the interactive part of a page less well than the static page (Xiao et al., 2024).

Reading-technique experiments in software inspection compare structured methods with unaided reading. In a randomized experiment with 48 graduate students inspecting requirements documents, reviewers using a scenario method, each searching for one class of faults, detected more faults than reviewers using ad hoc or checklist methods, and checklist reviewers were no more effective than ad hoc reviewers (Porter et al., 1995). The scenarios there were class-specific search techniques rather than walks through the steps of a usage scenario (Porter et al., 1995). The Maryland group reports about 35 percent more defects for defect-based readers than for ad hoc and checklist readers (Experimental Software Engineering Group, 1998). Perspective-based usability inspectors found more problems overall than heuristic evaluators, about 30 percent more when pooling three inspectors (24 professionals) (Zhang et al., 2000).

None of this tests states of an interface directly, and the inspection results did not always replicate. Defect-based reading was replicated independently in five contexts, with students and practitioners obtaining different results; the authors hypothesise that the technique is too sophisticated for undergraduates and that practitioners may revert to familiar techniques, so the gain depends on who reads (Basili et al., 1999). Two replications comparing perspective-based reading with a checklist gave conflicting results on one of the two documents, and each method uncovered defects the other did not (Fabbri et al., 2004). A meta-analysis of perspective-based reading experiments found no clear positive effect, as it beat ad hoc reading but did worse than checklists (Ciolkowski, 2009).

Review by someone who did not build the interface

Builders may miss problems that outsiders see. In a heuristic evaluation of a web prototype, the authors report relatively few differences between the developers who built it and user-interface experts, the experts finding 102 heuristic violations and the four developers 64 (Levi & Conrad, 1996). A practitioner column says confirmatory bias makes designers less likely to discard faulty designs, and also records practitioners who think the advantages of self-testing probably outweigh the bias; none of it cites a measured result (Sherman, 2009).

The case for independent review is largely practitioner opinion. What is measured is that independent evaluators also differ widely from each other: only 20% of 93 problems were found by all four evaluators analysing the same test sessions (Jacobsen et al., 1998), and a review of eleven studies found two-evaluator overlap from about 5 to 65 percent (Hertzum & Jacobsen, 2003). This evaluator effect is treated in Usability evaluation.

Agents in the loop

When an AI agent builds an interface for a user, the process questions are what it asks first and what it shows. In a within-subjects study of 12 people building interfaces with an AI chatbot, answering guided questions before generation led to fewer iterations and revisions than free-form prompting, at the cost of extra time (Agarwal et al., 2026). Seven of eight UI and UX experts preferred an agent that asks questions over one that passively follows instructions when building a web page from a sketch (Li et al., 2024b). In a 30-person task study (text and image generation, not interfaces), asking clarifying questions raised trust relative to no questions and also raised cognitive overhead, and timing made little difference (Ho et al., 2026). In a study with 16 professional designers, a structured specification of interface intent beat a conversational prompt tool on intent alignment (Chen et al., 2025b). A method that gives the model a crowdsourced dataset of user preferences (720 control preferences from 50 users) produced interfaces that users judged aligned with their preferences, tested with 72 users on six unseen tasks (Liu et al., 2026b).

How the AI is placed relative to the person matters as well. In a randomized experiment (N = 486), a mode in which the LLM rewrote people’s ideas raised quality but lowered diversity and perceived ownership, while reflective modes in which the LLM asked or suggested kept both (Maier et al., 2025). Letting app users pick from generated alternative interface improvements produced feedback judged better by users and developers than plain comments (Wei et al., 2026).

Most of these are preprints with small samples, and several do not concern interfaces. Using agents or simulated users to evaluate the result is the subject of Usability evaluation; what such agents produce is described in AI-generated interfaces.

Expressing visual preferences

People often cannot put a look into words. Experienced designers report that they and their clients use visual references more than words to discuss designs, and a tool showing dimensions drawn from generated images helped 52 novices write longer, more specific prompts (Tao et al., 2025). A practitioner analysis of public image-generation prompts proposes that users can see an image in mind but lack the words, and suggests style thumbnails; it reports no measured result (Mugunthan, 2023). Models fitted to one designer’s own annotations predicted that designer’s preferences better than a model trained on majority votes, and twenty professional designers comparing pairs of generated screens agreed only weakly on which was better (Peng et al., 2025). Visual appeal of websites differs by group (gender, education, country), so one person’s or one group’s picked examples do not give a universal target (Reinecke & Gajos, 2014).

Whether pictures beat words is not settled. A literature review reports that some conjoint studies of verbal versus visual stimuli found little difference in preference models while others found non-convergent results, and, citing Vriens (1995), that for attributes about design visual stimuli are needed for unbiased results (Berneburg, 2007). In shopping experiments, people preferred pictures to text, but with 27 options verbal descriptions gave better recall and images caused choice overload (product choice, not design taste) (Knowledge at Wharton Staff, 2013). A design-education paper states, citing earlier work, that highly figurative images in a mood board can block new concepts and that abstract imagery often works better, with no measured test (Munk et al., 2020).

Questions, controversies, and new developments

Does parallel prototyping pay for its cost?

The benefit of parallel prototyping varies with the task and costs time. The first lab’s second experiment (an egg-drop task, 35 participants) found no performance difference between parallel and serial, and the authors attribute this to task differences (Dow et al., 2009). Nielsen and Faber state that parallel design wastes effort, with several designers doing the same work, and that iterative design is preferable when cost matters more than early release (Nielsen & Faber, 1996).

People also resist it. Earlier work from the Murphy and Linsey lab found participants strongly preferred iterative prototyping even though parallel gave better designs (Sjolund et al., 2025). In the independent competition study, very few participants scored and participants believed iteration to be more effective (Murphy et al., 2022). When first-year students could choose their own strategy, the condition taught parallel prototyping scored higher in the competition, but none of those students actually used it, so the advantage is not explained by the method (Sjolund et al., 2025).

Do examples and AI output narrow the search?

Seeing others’ designs may reduce the range of ideas explored. Perttula and Liikkanen found that students shown four example designs surveyed fewer solution categories than controls but did not conform more to the examples’ categories, and attribute the narrower search to examples pre-exhausting the common ideas, not to fixation (Perttula & Liikkanen, 2006). In a 60-person experiment, support from an AI image generator during ideation led to higher fixation on an initial example and to fewer, less varied and less original ideas than a baseline (Wadinambiarachchi et al., 2024). In a study of AI-generated drone designs, participants who saw only numerical performance data picked the best designs best, and the way designs were shown changed which were chosen (Xu et al., 2025a).

How many alternatives to show is also open. A meta-analysis of 63 conditions from 50 experiments found a mean choice-overload effect of virtually zero, with considerable variance between studies and no sufficient conditions that reliably cause it (consumer choices) (Scheibehenne et al., 2010), so it supports no firm cap.

How much should an agent ask?

Questions before building carry costs. LHAW names three: interruption, latency and loss of trust from excessive questions. In its benchmark models mostly did not recover their original performance after asking clarifying questions, and one model over-clarified in 38% of trials (Pu et al., 2026). In a user study of a coding agent trained to ask questions, participants criticised it for excessive or generic questioning and for looping when details were unavailable (Sun et al., 2025). In a software-engineering clarification study more questions did not reliably help while sets with a higher share of answerable questions performed better (Vijayvargiya et al., 2026).

Timing and quality matter more than count. In a long-horizon agent benchmark, asking frequency did not predict task success, and clarification given after about 10% of execution (goals) lost nearly all value (Gulati et al., 2026). In an 89-participant web-search study, low-quality clarifying questions disturbed users and showing every clarifying question was risky, while showing only high-quality ones gave better gains with less effort (Zou et al., 2022). Limiting an agent to one to three questions cut user-attention cost by 36 to 73 percent while keeping 86 to 94 percent of the task reward (Zhang et al., 2026a).

Broader connections

The steps here feed the rules of layout and perception described in Interface design, and the output of agents that follow them is examined in AI-generated interfaces. Shared components and tokens that constrain what a designer or agent produces belong to Design systems. Which findings from an evaluation to act on, and how many users or evaluators to use, is the subject of Usability evaluation.

Further reading

• ISO 9241-210:2010 contents pages (International Organization for Standardization, 2010): the four design activities and the iterative framing.
• Dow et al. (2009) (Dow et al., 2009): the web-advertisement experiment behind most claims for parallel prototyping, with its egg-drop second experiment that found no difference.
• Tohidi et al. (2006) (Tohidi et al., 2006): users rate a design more critically when it is shown among alternatives, and give few constructive suggestions.
• Virzi et al. (1996) (Virzi et al., 1996): low- and high-fidelity prototypes find largely the same problems.
• Porter, Votta and Basili (1995) (Porter et al., 1995): the scenario-versus-checklist inspection experiment that later replications qualified.
• Ayoola et al. (2024) (Ayoola et al., 2024): a randomized test of what personas change in what developers prioritise.