> ## Documentation Index > Fetch the complete documentation index at: https://docs.leonardo-impact.com/llms.txt > Use this file to discover all available pages before exploring further. # How voi writes your interview > The principles voi follows when it turns your project into an interview guide, and how it conducts that interview with each respondent. voi builds every interview in the same sequence. It works out what your project needs to measure, chooses or writes the questions, puts them in order, and then conducts the interview with each respondent. This page explains the principles behind each step, so you can judge the guide voi gives you and explain it to others. For the controls themselves, see [Plan](https://docs.leonardo-impact.com/using-voi/create.md). > [!NOTE] > > The wording, ordering and interviewing principles on this page follow guidelines from voi's scientific partner, the Technical University of Munich (TUM). ## From your project to outcomes Everything starts from your project description and any documents you add. voi turns them into your activities, the groups of people you serve, your impact objectives, your study focus and the attributes you want to compare results by, such as age or location. Names and figures are taken from your sources as written. The study focus holds an optional purpose and any further topics you want to learn about beyond your Theory of Change. Your Theory of Change then works backwards from the impact you aim for. Each outcome describes one change for the people you serve, such as their income or their confidence, carries a short reason for how that change happens, and links to the impact objectives it serves. voi aims for a few outcomes per activity, so that each one can be measured well, and it keeps every outcome you or your documents name. It merges a long-term outcome that only restates a short-term one, so the interview asks about each topic once, and tells you whenever it trims or merges. Every outcome and every further topic in your study focus becomes something the interview has to cover. ## Choosing indicators For each outcome, voi looks for a validated indicator in its library and uses one only when it fits the core of the outcome strongly. It also weighs how well the indicator suits your study, such as the people you serve and your sector, and whether it asks more of respondents than your study needs. A partial match appears as a suggestion you can add yourself, and where nothing fits strongly, voi writes a question for that outcome instead. Each outcome gets either a library indicator or a written question, so no outcome is measured twice. A chosen indicator always comes with all of its questions, because its score needs every answer. For changes in knowledge, attitudes, confidence or wellbeing, voi prefers measures in which people rate their own experience over measures that count outside behaviour, when both fit equally well. When you regenerate the guide, a question you have locked keeps its place, and voi adds nothing else for the outcome it covers. ## How questions are worded When voi writes a question, whether it composes the whole guide or you use **Add question**, **Rephrase with AI** or **Regenerate with AI**, it follows the same principles: * Each question asks about one thing. "Did your harvest change in quantity or quality?" becomes two questions. * The wording never suggests an answer. voi asks "How has the programme affected you, if at all?" instead of "How much has the programme helped you?" * Questions use plain, everyday words, without research terms such as "beneficiary" or "intervention". * A question about change names its reference point, such as "Compared to before you joined" or "In the past six months". * Open questions invite a story and start with words such as "how" or "what". They ask how something came about instead of why, so respondents describe what happened without having to defend a choice. * Within one outcome, what went well comes before difficulties, and what happened comes before how it felt. A question about a difficulty points to circumstances, never to the respondent's shortcomings. Answer options get the same care. Scales are balanced, with as many options on one side as on the other, the options of one question read alike so the expected answer never stands out, number ranges follow on without gaps or overlaps, and every option names one specific answer. A question about how someone sees or judges something stays open or uses a scale, and never asks for a plain yes or no. Questions from a library indicator keep their validated wording, because changing it would break comparison with other studies that use the same indicator. ## Automatic checks voi checks the questions it writes, and the ones you write yourself, for problems a rule can catch reliably. The checks look at the wording (a question that asks about two things at once, or opens like a yes or no question), at the answer options (ranges that overlap or leave a gap, options that differ in form), and at the introduction. While voi composes the guide, a question that fails a check goes back to the AI with the reason, so most problems are fixed before you see the guide. When a note remains, it appears under the question in the interview outline, on suggestions in the **Add question** dialog, and under the fields of the question sidebar while you edit. Each note says what it found and how to fix it, and respondents never see it. Notes are advice, and only a few problems, such as a closed question with fewer than two options, stop you from saving. ## How the interview is ordered The interview outline has three question sections: | Section | What it holds | | - | - | | Warm-up | An easy, open opening question, then the questions about who is answering | | Core | The Net Impact Score questions if you chose them, then the questions for your outcomes, in the order of your Theory of Change, and for your study focus, then the Most Significant Change questions and the Net Promoter Score question if you chose them | | Closing | One open question that invites anything important the interview has not asked | A question that decides whether another one appears always comes before it. Without a target length, voi covers every outcome first and keeps the interview as short as that allows. With a target, voi treats it as a cap: it takes out the questions that matter least to your outcomes first, keeps the questions of an indicator together because its score needs every answer, and removes the last question of an outcome only when nothing else fits, telling you which outcome did not fit and how you could make room. Questions you locked, questions you wrote yourself, the questions about who is answering and the questions of the commonly chosen indicators you picked always stay. See [Setting a target completion time](https://docs.leonardo-impact.com/using-voi/create.md#setting-a-target-completion-time). ## The introduction The introduction is the first thing a respondent hears from voi. It always opens with the same sentence, "Hello! My name is voi, and I'm an AI researcher here to listen to your experience." It then says briefly what the interview is about and that there are no right or wrong answers. It makes no promise your study may be unable to keep, such as anonymity, payment or an exact duration, and it avoids phrases that assume a positive effect, such as "how we helped you". If a draft breaks these rules, voi uses a standard version instead. You can edit the introduction in the outline, and a note appears if your edit adds leading language. The consent text on the welcome screen is separate. It comes from a template you can customise, and you stay responsible for what it says. See [Share](https://docs.leonardo-impact.com/using-voi/share.md#the-consent-and-welcome-screen). ## During the interview voi asks one question at a time, exactly as written, and shows the answer options on screen instead of reading them out. When it introduces a scale, such as one of agreement or frequency, it names the first and last options exactly as the respondent sees them, for example from "Strongly disagree" to "Strongly agree". After an answer, it may ask follow-up questions, within the range set for that question: * A follow-up picks up something specific the respondent just said. voi never repeats or summarises the answer back. * It asks first what happened and later how it felt, and it asks how something came about instead of why. * An answer at the very end of a scale, such as "Much worse", gets a follow-up about what actually happened, when follow-ups are on for the study. voi stays warm and non-judgemental and never pushes for more than a respondent wants to share. It does not probe distress or a sensitive disclosure, and when distress is apparent it tells the respondent that they can skip a question or stop. If a respondent says they were paid, told or pressured to answer a certain way, voi acknowledges it once, and the answer carries a label you can see in [Review](https://docs.leonardo-impact.com/using-voi/review.md). To see these principles at work on your own study, open the interview outline in [Plan](https://docs.leonardo-impact.com/using-voi/create.md). For how a single question is built, see [Writing questions](https://docs.leonardo-impact.com/building-indicators/writing-questions.md). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.