From prompt engineering to dialogue

From Prompt Engineering to Dialogue: A Different Way of Working with AI

When ChatGPT first became widely available in late 2022, we were told that we needed to learn prompt engineering. The quality of AI-generated responses seemed to depend on how precisely we formulated our instructions. We learned to assign roles, specify formats, provide examples, break tasks into steps, and carefully structure our prompts.

Entire collections of prompt templates emerged. Some were remarkably elaborate, instructing the AI to act as an experienced qualitative researcher, apply a particular methodology, and follow a predefined sequence of analytical steps.

Much of this advice made sense at the time. Early language models were less capable of following complex instructions, had limited context windows, and often needed considerable guidance to produce useful results. But the technology has moved on.

What has changed? Today's language models are much better at understanding natural language, following instructions, maintaining conversational context, and responding to follow-up questions.

This does not mean that prompting has become irrelevant. Clear instructions, appropriate context, and examples still matter, particularly when we need consistent outputs or want the AI to perform a specific task. What has become less necessary is the idea that every interaction requires a carefully engineered prompt.

There is also an important technical distinction here. When we ask a question in ordinary language, the AI does not first translate it into some perfectly optimised prompt behind the scenes. It processes our language directly, together with the available conversational context. Modern models have simply become much better at doing this.

For qualitative researchers, this development opens up an important possibility: We can move from instructing AI to perform analytical tasks towards engaging with our data through an ongoing dialogue.

From an analytical task to an analytical conversation Consider the difference between these two ways of working. A conventional prompt might say:

Analyse the following ten interviews and identify five recurring themes. Provide a description of each theme and supporting quotations.

This is a perfectly reasonable instruction. We ask the AI to perform a task and produce an output. We can then evaluate the result. But qualitative analysis rarely proceeds in such a straightforward manner. Our understanding develops as we engage with the material. One observation raises another question. An unexpected contradiction changes our interpretation. We recognise something that we had not considered before. Instead of asking AI to deliver a finished analysis, we can use it to support this developing process.

We might begin with a simple descriptive question and gradually move towards more conceptual and interpretative questions. For example, imagine we have conducted ten interviews about people's experiences of working from home.

1. Descriptive questions — understanding the material • What do participants say about the advantages and disadvantages of working from home? • Which aspects of their everyday routines do they describe? • What different experiences of social isolation appear in the interviews? At this stage, we are exploring what participants have said and becoming familiar with the material.

2. Comparative questions — exploring differences and relationships • How do experiences differ between participants who live alone and those who live with others? • Are there differences between participants who value flexibility and those who prefer a structured working day? • Do any participants describe experiences that contradict the patterns identified so far? Now we begin examining variations, relationships, and exceptions rather than simply collecting recurring topics.

3. Conceptual questions — moving towards interpretation • Could the tension between flexibility and structure be understood as a question of control over one's working life? • How is the boundary between professional and private life negotiated? • What might explain why the same working arrangement is experienced as freedom by some participants and as a burden by others? Here, the researcher begins introducing analytical ideas and tentative interpretations. These are not findings simply because the AI can formulate a convincing response. They are propositions to investigate against the empirical material.

4. Synthesising questions — developing an understanding • Considering our discussion so far, what relationships are emerging between autonomy, social connectedness, and work satisfaction? • How might these relationships help us explain the differences between participants? • Which observations support our emerging interpretation, and which remain difficult to explain? At this point, the conversation can help us bring together different analytical observations and develop a more integrated understanding.

The important difference is not the wording of the question The examples above could all be written as individual prompts. What makes the approach dialogic is not simply that we use natural language. It is that each question can build on what we have learned from the previous exchange.

We can challenge an AI-generated interpretation, introduce our own ideas, request contradictory evidence, return to a particular interview, or reconsider an earlier assumption. The analytical direction is not determined entirely in advance.

This is where I see particular value for qualitative research. Rather than automating a predefined sequence of methodological procedures, we can use AI as a conversational instrument within an evolving process of inquiry.

Of course, a language model does not understand our research in the same way we do. It can overlook context, overgeneralise patterns, or produce interpretations that sound plausible but are insufficiently grounded. Even with today's larger context windows, maintaining reliable connections to the original material remains essential.

This is why the researcher needs to remain actively involved: examining the evidence, questioning interpretations, bringing contextual and theoretical knowledge into the conversation, and deciding which analytical directions are worth pursuing.

Moving from prompting to analytical dialogue opens a different way of conducting qualitative inquiry, one in which communication itself becomes a central means of developing analytical understanding. This was my vision when I realized the potential of this new technology early on on coined the sentence: Stop Coding. Start Talking!

In 2023, the potential was already visible, but the technology still imposed considerable limitations. Today, with larger context windows, stronger language understanding, and increasingly natural spoken interaction allows this vision to become more real, step-by-step. What many researchers still need to learn is how to interact with an LLM in natural language and not treating it as an answering machine.