The flipped interaction pattern is a prompting technique in which the user instructs a language model to ask the questions instead of answering them, gathering information step by step until it can accomplish a stated goal. Rather than writing one long prompt that anticipates every detail, the user defines the objective, the stopping condition and the question style, and the model conducts a structured interview.
Prompting research catalogs it among the interaction-category prompt patterns, alongside patterns that shape how a conversation unfolds rather than what a single answer contains. Flipped interaction, sometimes simply called flipped interaction, is most useful when the user knows the goal but does not know which details matter, which is common in requirements gathering, diagnostics, tutoring and personalized planning.
How the flipped interaction pattern works
The pattern works by reversing the default roles of a chat session: the model becomes the interviewer and the user becomes the source of facts. A typical instruction reads: “Ask me questions, one at a time, to collect what you need to design a weekly training plan. When you have enough information, produce the plan.” The model then drives the dialogue until the stopping condition is met.
A well-formed flipped prompt contains a small set of components. Each one removes a specific source of ambiguity, and omitting any of them is the usual reason the pattern fails.
| Component | What it specifies | Example wording |
|---|---|---|
| Goal | The deliverable the questions serve | Draft a data retention policy |
| Questioning rule | Number of questions per turn | Ask one question at a time |
| Stopping condition | When to stop asking | Stop after you have enough, or after 8 questions |
| Scope limits | Topics to cover or avoid | Ask only about legal and storage constraints |
| Output format | Shape of the final result | Return a numbered policy with headings |
The mechanism relies on the model’s ability to identify missing variables for a task. A model that has seen many examples of travel itineraries, for instance, can infer that budget, dates, mobility constraints and interests are needed. The pattern turns this latent task knowledge into explicit questions instead of silent assumptions.
Variants and design choices
Flipped interaction has several practical variants that differ in pacing, structure and how much control the user keeps. The right choice depends on how tolerant the user is of long exchanges and how precise the final output must be.
| Variant | How it behaves | Best suited for |
|---|---|---|
| Single-question turns | One question, then waits | Careful, high-stakes intake |
| Batched questions | Three to five questions per turn | Busy users and simple tasks |
| Adaptive branching | Next question depends on prior answers | Diagnostics and troubleshooting |
| Capped interview | Fixed maximum number of questions | Time-boxed sessions |
| Confirm-then-generate | Summarizes facts before producing output | Specifications and contracts |
A confirmation step before generation is especially valuable. When the model restates the collected facts and asks the user to correct them, misunderstandings surface before they reach the final output, which is far cheaper than fixing a long generated document afterward.
Comparison with related prompting techniques
Flipped interaction differs from neighboring techniques in who supplies the context and when. Several related patterns also involve questions, but they direct the questions toward different parties or purposes.
| Technique | Who asks questions | Main purpose |
|---|---|---|
| Flipped interaction | Model asks the user | Collect missing requirements |
| Socratic prompting | Model asks the user | Guide reasoning or learning |
| Self-ask prompting | Model asks itself | Decompose a multi-hop question |
| Iterative prompting | User refines the prompt | Improve output over rounds |
| Role prompting | Nobody by default | Set expertise and tone |
The patterns combine well. Pairing flipped interaction with role prompting, for example “act as an experienced tax adviser and interview me”, improves the relevance of the questions, while a prompt template can store the whole interview structure for reuse across a team.
When to use the pattern and how to implement it
The pattern pays off when the information gap sits with the user, not with the model. It is less useful when the user already holds a complete specification, because the interview then adds turns without adding facts.
| Situation | Recommendation | Reason |
|---|---|---|
| Vague goal, unknown details | Use flipped interaction | Questions expose hidden requirements |
| Complete written brief | Use a direct prompt | No missing information to collect |
| Automated pipeline | Avoid the pattern | No human available to answer |
| Novice user | Use single-question turns | Reduces cognitive load |
| Expert user in a hurry | Use batched questions | Fewer conversational turns |
In applications, the pattern is often placed in the system prompt or custom instructions, so that every session begins with an interview. Developers commonly track collected fields in structured state, such as a JSON object, so the application can verify completeness instead of trusting the model’s judgment alone.
Limitations and common mistakes
The main weakness of flipped interaction is that the model decides what matters. If its picture of the task is incomplete, it will ask confident questions about the wrong things and skip a critical constraint. Several recurring mistakes follow from this.
- No stopping condition, which lets the interview continue indefinitely or end abruptly.
- Too many questions per turn, which overwhelms users and produces shallow answers.
- Leading questions that steer the user toward the model’s default assumptions.
- Skipping the confirmation summary, so misread answers flow straight into the output.
- Collecting sensitive personal data without a clear need, which creates privacy risk.
Output quality also remains bounded by user knowledge. If the user cannot answer a question accurately, the model may fill the gap with plausible defaults, so critical outputs still need expert review.
A non-technical view
Usually, people ask an AI chatbot a question and get an answer. The flipped interaction pattern turns this around. You tell the chatbot your goal and ask it to interview you first.
Think of a good tailor. A tailor does not guess your size. The tailor takes measurements, asks about the occasion and only then cuts the fabric. In the same way, the chatbot asks about the details it needs. The result fits your situation instead of a generic average.
A concrete example
A small accounting firm wants a written onboarding checklist for new clients, but the office manager is unsure which details belong in it. She writes one flipped prompt: act as an experienced practice manager, ask me one question at a time about our services and processes, stop after at most 12 questions, summarize what you learned for my confirmation, and then produce a checklist grouped by week.
The model asks about service lines, document formats, filing deadlines, identity verification and software access. After 10 questions it presents a summary, and the manager corrects one point about how tax documents arrive. The final checklist covers four weeks with 23 items. The first draft needs only minor edits, whereas an earlier one-shot attempt had missed identity checks entirely and assumed paper records the firm no longer uses.