AI as a reasoning partner: the role of prompting in UX

By Arianna F. Scopece, UX/UI Designer at Excellence Innovation

In recent years, the spread of Generative AI has profoundly transformed the way we design digital products and services.
Tools such as ChatGPT and other automatic generation platforms have enabled us to produce content, analyse information, summarise documents and prototype solutions in extremely short timeframes.
In an initial phase, use mainly focused on the output, treating AI as an execution tool.

Today, however, the goal is no longer only to obtain rapid results, but to integrate AI into processes in a structured, reliable and coherent way. It is in this transition that a paradigm shift emerges: AI becomes a reasoning partner, and prompting establishes itself as one of the key skills of UX. Interacting with these systems means going beyond the input-output model and introducing logics of dialogue, iteration and control.

From prompt control to the interaction process

Anyone can write a prompt: advanced technical skills are not required.
However, when AI begins to be integrated into design processes, a clear need quickly emerges: to obtain control and define specifications that are as precise as possible.

In the UX field, a widespread approach consists of building structured prompts, similar to actual project briefs:

• overview and context
• main actors and stakeholders
• target users
• functional specifications
• UX/UI and accessibility principles
• use cases and user stories
• objectives and expected results
• technical constraints and open points

The expectation is clear: in response to a well-defined input, the output should be coherent, replicable and of quality.

In practice, however, some critical issues quickly emerge. Even when using the same prompt, the generated outputs can vary significantly: not only across different tools, but also within the same tool, when repeating the request in different sessions. This results in:

• differences in the outputs, both in structure and in the interface
• cases not considered
• loss of important elements
• difficulty in managing complex projects

These aspects highlight a fundamental characteristic: even with structured inputs, generative AIs do not produce fully predictable outputs. The focus therefore shifts from the single request to the process of interaction with AI.

The limits of a structured prompt

When something does not work, there is a tendency to increase the level of detail, because it is more immediate than adopting new strategies or techniques.

We therefore structured more precise formats, using languages such as XML or Markdown to organise information rigorously and guide AI in the generation of components or interfaces. In particular, by leveraging Figma Make, where the precision of input data is fundamental.

Even in this case, new limits emerge:

• very long and complex prompts, with the need to split them into several parts
• longer generation times
• constraints related to tool credits or performance

And, above all, a counterintuitive aspect: even by increasing the level of detail, some information continues to be lost. The fragmentation of the prompt, which is necessary to manage complexity, in fact introduces new discontinuities into the process.

This highlights how the quality of the result does not depend only on the precision of the input, but on the way in which the entire flow of interaction with AI is built and managed.

The role of doubt: designing for error as well

Another central element concerns the role of doubt in the process.

Interaction with AI tends, by its nature, to be confirmatory: the system returns answers that are coherent with the input received, rarely questioning the initial choices.

For this reason, it becomes fundamental to intentionally introduce verification mechanisms:

• questioning one’s own prompt
• reflecting on the clarity of the requests
• explicitly asking AI to highlight limits and possible errors
• never assuming the output to be correct by default

This approach directly recalls UX practices related to validation and testing: one does not design only the solution, but also the way in which it is verified.

Conversational UX: from prompt to designed dialogue

All these limits have led to a change in perspective, pushing us to question the initial approach. At the same time, the difficulty of fully controlling AI, combined with the speed with which these technologies evolve, makes the need to rethink the way we interact with these systems even more evident.

To this, a further element is added: the market has produced an enormous quantity of AI-based tools, without leaving the necessary time to truly understand them and integrate them into processes. The limit, therefore, lies not so much in the capabilities of the technology, but in the time and space we have to learn it and use it consciously.

If prompt structuring is not sufficient to guarantee quality and each tool interprets its own language, the focus shifts from the single input to the interaction process. It is no longer a matter of insisting on a single model or providing all the information at one single moment, but of building the result through dialogue.

A particularly effective pattern consists of introducing a co-reasoning phase: before generating the final output, the system is guided to ask questions, challenge the initial assumptions, highlight ambiguities and signal possible critical issues.

This key step allows us to:

• make the process more transparent, controllable and, why not, more human
• bring out missing information or edge cases not considered
• explore design alternatives or validate a solution
• guide AI in the client’s style, patterns and UI

From a UX point of view, this means designing an interaction that includes feedback and iteration, instead of limiting oneself to an input-output logic.

Conclusion

In conclusion, the integration of AI into design processes does not eliminate the need for design skills, but makes them even more central. In complex contexts, in fact, such as B2B contexts and in particular in the private banking sector, the use of AI introduces additional levels of criticality: highly technical content, stringent requirements, specialised users and a strong need for consistency and reliability.

In these scenarios, AI can effectively support content generation, the exploration of design alternatives and scenario analysis. However, it is not able to replace the understanding of context, the management of complexity and the qualitative assessment of solutions. The human component therefore remains central in interpreting information, identifying errors or inconsistencies and making conscious design decisions.

It is precisely in this space that the role of the UX designer is strengthened, responsible not only for the final product, but also for the quality of the interaction with the AI system.

The experience gained in the use of AI leads to a key awareness: there is no perfect prompt, but there is an effective way to design the conversation.

The value does not lie in the single request or in the generated output, but in the ability to build a dialogue, reason and iterate consciously, introduce control checks and integrate critical thinking throughout the entire process.

In this scenario, AI ceases to be a simple operational accelerator and becomes our reasoning partner with whom to co-design solutions. A collaboration that, in order to be effective, requires method, experience and a solid UX culture.