From the first idea to a better question

A question such as “How can we use AI so that customers receive an answer more quickly?” sounds like a clear assignment. Yet it already contains a diagnosis and a solution: response time is the problem, AI is the appropriate change, and “an answer” can be generated automatically. That may be true. The bottleneck may instead be missing information, unclear responsibilities, or an answer that is possible only after a check. Anyone who writes a prompt immediately can process an unstated assumption very efficiently.

This article therefore does not teach how to collect as many questioning techniques as possible. It practises a narrower skill: in Module A of the project-developed questioning model F v5, choosing precisely the supporting technique that fits an identifiable gap. The same case question then visibly runs through F v5 in the order A→B→D→C. A clarifies and chooses; B opens up several explanations; D takes up follow-up questions, evidence gaps, fallback, and revision; C consolidates the revised question, the transparent path, and the bounded next step. A single technique can support one part of this work. It does not replace the model sequence and does not establish an answer.

The guiding question is: How does a learner or small team select an appropriate questioning support for an unclear AI task without handing the work over to that technique?

A question is already an intervention

An initial question sorts reality in advance. It determines what appears to be the problem, which measure counts, and who is treated as responsible. “How do we automate customer service?” makes automation the goal. “Why does an answer often take so long?” centres duration. “Which kind of request can be answered reliably, and what information is missing for it?” opens a different inquiry. None of these questions is neutral: each directs attention and distributes the burden of justification.

This matters for AI ethics because a neatly worded assignment does not guarantee a fair or well-informed decision. If only handling time is measured, additional correction work for employees can remain invisible. If only management’s perspective appears, the question of how to handle unclear or urgent requests may be absent. And if a case is already described as “text production,” a non-AI option such as a clear status page may never be examined.

The pragmatist philosopher John Dewey describes reflective thought as an inquiry that often begins with confusion, uncertainty, or doubt. An initially plausible explanation is not simply adopted; further observations should support or challenge it. In that sense, the better question is not merely a stylistic improvement. It is a provisional judgement about which situation should be investigated and which observation could change the explanation. Dewey’s work is a philosophical lens here, not a four-step instruction and not empirical evidence of F v5’s effectiveness.[1][2]

Plato’s Socratic dialogues also help distinguish an example from a definition. In Euthyphro, Euthyphro first gives a concrete action as an example of piety; Socrates asks for the characteristic shared by all pious actions. A later definition is tested and does not lead to final agreement.[3][4] Such questioning can make conceptual boundaries and unresolved questions visible. For a learning exercise, this gives a useful stance: a clarifying question may interrupt a premature yes or no. It must not be used to examine a person or as a rhetorical trick.

The philosophical insight is practical: sometimes a question’s most important achievement is to loosen an undeserved certainty. A good result may then be: “We do not yet know whether the bottleneck is writing or access to information.” That is not inaction. It identifies the observation missing before a tool decision.

F v5 is the sequence; supporting techniques are bounded aids

The project-developed model F v5 explicitly orders question expansion as A→B→D→C. The order may look unusual, and it matters: D covers error handling and iteration and comes before C, consolidation. The final question and adapted working prompt are therefore formulated only after uncertainties, feedback, and needed revisions have been addressed.

A – clarify input, context, terms, goal, and assumptions. First, make visible what the input says and what it already assumes. A supporting method is added only when it fulfils a concrete subtask: clarifying an ambiguous term, investigating a local cause as a hypothesis, or varying a premature solution idea.

B – develop several explanations and meta-questions. The starting point produces competing hypotheses or questions. B asks not only “What might be true?” but also “What assumption is contained in this explanation?”, “Which alternative is also plausible?”, and “What feedback would change our focus?” Feedback from users or affected people can shift the work’s focus; it is not merely counted as agreement.

D – address uncertainty, fallback, and revision. Missing information leads to a follow-up question or a clearly bounded fallback. Suitable external data, best practices, risks, and industry context can be named as items to check. A revision must make visible what new information changed. D is not a retrospective record placed after an answer that has already been fixed.

C – integrate and hand over results transparently. Only here is an adapted final question or working prompt formulated. A short process summary records which terms were clarified, which hypotheses were considered, which feedback was incorporated, and which limits remain open. The GROW reference helps turn the question into a bounded next step.

The three supplied primary pages on F v5 describe this model architecture and its intended operations. They document a project-developed methodological design; they are not a published effectiveness study. Nor does the origin of a supporting method establish that its combination with F v5 has been scientifically validated. The article therefore distinguishes between model description, philosophical framing, supportable descriptions of individual methods, and open effectiveness questions.

QFT is neither applied in this article nor used as its comparative framework. The primary F text lists QFT as one possible support in A. The learning purpose selected here requires no additional question-generation routine. The article therefore develops neither QFT steps nor QFT materials or a QFT exercise sequence.

A starts with a selection question: what gap do we have?

Anyone who runs five or six methods one after another for every task has not yet worked methodically. A method buffet can create an impression of thoroughness even though the central question remains unchanged. Selection should instead follow the observable gap:

The table is a selection aid, not a prescribed stack of methods. In the 30-minute format, exactly one A support is chosen. A second support is permitted only in a new F pass if an additional, different kind of gap has become visible; it then needs its own reason for selection and its own A result. It is not permissible to tick off several supports as a precaution. If a support does not change the result, it does not belong in the sequence.

Clarifying terms: Socratic questions

A Socratic clarifying question fits when the initial question uses an apparently self-evident word. “Fast,” “safe,” “fair,” “automatic,” or “answer” can mean different things. For a customer, “fast” may mean a short acknowledgement; for the team, it may mean a final, factually reliable response. These meanings are not interchangeable.

The working movement is: name the term, ask for its meaning in the concrete case, and check a criterion and counterexample. “What counts as an answer here?” can show that an acknowledgement and a reliable delivery update are two different services. “How would we recognise that an answer was helpful?” separates the organisation’s perspective from that of the person who must act on the information.

The Socratic model should not be treated as a fixed question form. Historical and philosophical interpretations of what exactly “the Socratic method” is are not uniform. For this article, its bounded function is enough: expose term and criterion, then check whether they fit the examples. Someone who asks questions in quick succession without waiting for an answer or being willing to revise their own interpretation is not clarifying together but interrogating.

Investigating causes: Five Whys

Five Whys fits when a concrete observed process or error occurs in a manageable process chain. The next why-question takes the previous answer as a provisional starting point. In the Toyota and Lean tradition, repeated questioning is used to look beyond a visible symptom for conditions in a concrete process problem.[5][6]

The phrase “ask five times” must not become a proof ritual. The last answer is not an automatically discovered root cause. Several causes may operate in organisations at once: missing information, changing responsibilities, seasonal workload, or an unsuitable success criterion. F v5 therefore records the answer as a hypothesis. B opens competing explanations; D asks which process observation or source could support or challenge them.

Five Whys does not fit well when the actual problem is a disputed value, an unclear term, or an open future option. “Why should AI be used?” is not a linear error analysis. Here the technique risks retrospectively securing a direction that has already been chosen through ever deeper reasons.

Varying ideas: SCAMPER

SCAMPER directs seven kinds of change question at an existing idea or object: substitute, combine, adapt, modify, put to another use, eliminate, and reverse or rearrange. Robert F. Eberle presented the letter sequence in 1971 as games for developing imagination; his 1972 article notes that the letters draw on an idea checklist by Alex Osborn.[7][8] Its use in A is narrowly defined: a question that knows only “Which AI tool?” can be opened to alternatives.

In the fictional customer-service case, “answer automatically” could be replaced or bounded by a clearer status page, a prioritised queue, an internally prepared response, or an automatic acknowledgement. SCAMPER provides search directions, not evaluations. The next questions are: Which variation fulfils the goal? What information does it require? Who must control it? What follows from a wrong answer? A creative option without a subsequent check remains an idea.

Ordering goals and the situation: GROW

GROW structures a conversation around Goal, Reality, Options, and Will: the desired goal, current situation, possible options, and next step. The model was developed by John Whitmore with colleagues and became widely known chiefly through his 1992 book Coaching for Performance.[9][10] Within F v5, GROW can help in A as a short situation check when someone presses to act before goal and starting position are clear. In C, the same reference can structure the bounded next step.

The dual use needs a boundary. In A, “Will” does not mean the team has already committed to an introduction. If a binding measure appears in A before B opens hypotheses and D addresses evidence gaps, the project-developed model’s order has been skipped. The goal itself may also be examined. GROW orders an intention; it cannot decide whether that intention is justified or whose interests it considers.

Making learning actions concrete: Bloom

The original Bloom taxonomy was developed as a classification of educational objectives in the cognitive domain. Its six main categories are Knowledge, Comprehension, Application, Analysis, Synthesis, and Evaluation.[11][12] The widely used verb sequence Remember, Understand, Apply, Analyse, Evaluate, and Create belongs to the 2001 revision; it is not the category list of the 1956 edition.[12] For this article, Bloom is not a questioning technique for arbitrary factual problems. The taxonomy helps a teacher name the expected learning action concretely.

Instead of “Learners understand AI ethics,” the goal might be: “Learners can distinguish an assumed solution, missing information, and a checkable criterion in an AI request.” Another task would be to compare two plausible explanations or revise a question using new evidence. Bloom makes such performance requirements more visible. A more complex activity is not automatically more correct, fairer, or more demanding for every group. The taxonomy is an orientation for objectives and tasks, not a staircase to morally better judgement.

The same case through A→B→D→C

For the exercise case, take a fully fictional small trading company. Customer service repeatedly receives questions about delivery status and availability. Management is considering having a language model answer such emails automatically so that customers are informed more quickly. There is no real company, no real customer data, and no measured response time.

A: choose one appropriate support instead of a method buffet

The initial question is: “How can we use AI so that customers receive an answer more quickly?” The bottleneck is initially conceptual: “more quickly” and “answer” remain unclear. The team therefore chooses Socratic clarifying questions as exactly one A support. The choice is reasoned: it is not yet about why a particular machine or process stage fails; nor has management asked to creatively vary an already known solution idea. First, the intended result must be clarified.

The follow-up question reveals two possible aims. First, customers may need to learn as quickly as possible that their message has arrived. Second, they may need a reliable update on actual delivery status as quickly as possible. An automatically generated acknowledgement fulfils the first aim, but not the second. A reliable status update may require current information from another part of the business. “Response time” therefore becomes two different measures: time to acknowledgement and time to a factually checked update.

The A result is not a new product decision. It is a more precise initial question: “Which recurring requests are waiting for checkable information, and which of them can be handled with a reliable status source?” The team also records the assumption that writing the email causes the delay. This assumption remains open.

B: let explanations compete

B develops several hypotheses instead of immediately treating one as a finding:

The hypotheses are proposals for checking, not statistical statements. A small business can first describe its own processes using a limited, privacy-conscious example. It does not need an AI system immediately. A simple category sheet can record whether a request needs an acknowledgement, a status update, product information, or other handling. A non-AI option would be a well-maintained status page or a clear internal route of responsibility.

A meta-question changes the view: who judges that an answer is “fast enough”? Management, which looks at throughput, or also the person who receives a wrong status update and must ask again? This does not impose an outside value on the case. It makes visible that time as a measure does not represent every consequence.