An interview is the guide your colleague answers: a title, the goals behind it, and the questions the AI interviewer asks. This article covers building one.
Start from a description
Go to Interviews → New Interview and describe what you want to capture in plain language. For example:
I want to capture what our departing operations lead knows about month-end close — the steps, the exceptions, and who she chases when numbers do not reconcile.
The assistant drafts a complete guide from that: a title, the goals, and a set of open-ended questions with sensible follow-up depth. You can also pick one of the starting points on the page, or choose Skip Guided Setup to write everything yourself.
Refine it in the builder
The builder has two panes. On the left you talk to the assistant; on the right you see the guide it is editing, and you can change any field by hand.
Ask for changes in plain language — "make question 3 about exceptions rather than the happy path", "add two questions about the tools she uses", "shorten the welcome message". The assistant rewrites the guide and the right pane updates.
Edits save themselves: a text field saves when you leave it, and reordering, deleting or changing follow-up depth saves immediately. The header tells you where things stand.
The fields
- Title — what the interview is called, shown to your colleague.
- Interview goals — what you want to learn and why. The AI reads this when deciding which follow-up questions are worth asking, so it is worth being specific.
- Background for the AI interviewer — context about your company, team or process. Use it for things the interviewer should know but should not ask about.
- Welcome message — shown before your colleague starts. Say who it is for and roughly how long it takes.
- Sections — optional groupings for the questions. See below.
- Questions — see below.
- Thank-you message — shown when they finish.
Questions and follow-ups
Each question has a follow-up setting from None to Up to 3 follow-ups. This is how deep the AI may probe before moving on. After each answer it decides whether one more question would surface something real — a concrete example, an edge case, a step that got glossed over — or whether to move on. It never asks something that was already answered, and it never exceeds your limit.
Use follow-up instructions on a question to steer that probing: "push for a specific recent example", "find out who she escalates to", "ask what happens when the file is late". It is free text and grows as you write, so a paragraph of judgement is fine.
The AI also knows when to steer back. If your colleague changes the subject, asks the interviewer something, or would rather not answer, it acknowledges what they said and puts the question again in plainer words. Steering back does not count against your follow-up limit, and if someone declines twice it moves on rather than pressing.
Reorder questions with the arrows, and remove them with the trash icon.
What a question can and cannot be
Every question is open-ended and answered in your colleague's own words, spoken or typed. Nuggetz has no multiple choice, no rating scales, no date or number fields, and no branching. Everyone is asked the same questions in the same order.
Questions and messages are plain text. You cannot put an image, a file, a diagram or formatting into one, and your colleague has no way to upload or attach anything back. When the knowledge is visual, ask for it in words — "describe what the dashboard shows when it goes wrong" gets you the part that matters anyway.
A colleague picks one language before they start and keeps it to the end; there is no switching mid-interview, and Nuggetz does not guess their language from what they say. Translations are regenerated from the guide rather than edited by hand.
The builder assistant will tell you when Nuggetz does not do something rather than agree to it. If you are still unsure, ask us.
Languages
Interview in is chosen on the create page, before you describe what to capture: it is the language the assistant drafts the guide in, and it cannot be changed afterwards. Colleagues answer in it by default.
On a draft you can add up to five participant languages for a colleague to pick instead. Everything they meet is in the language they choose on the first screen: the questions, the AI's follow-ups, the voice that reads them out, and the transcription of what they say back.
Adding a language does not translate it — that happens when you launch, so your later edits cannot leave a translation out of date. Launching translates every language that needs it and shows you how far it has got; allow ten to thirty seconds a language. If one fails, the interview stays a draft and you can launch again.
After launch, editing the guide marks the affected translations Needs update. Choose Update translations to regenerate them; new colleagues get the new version, and anyone part-way through finishes the one they started.
To try a language before launching, translate it early with the icon on its chip, then pick it in the Preview pane.
Results in chooses what summaries, nuggets and the handbook are written in — one language for the whole interview, whatever language each colleague answered in, so a French answer and an Italian answer reconcile into one set. Quotes stay in the respondent's own words. It starts on the language the interview is in; change it if your team reads results in another.
Every language setting is fixed once the interview launches.
Sections
Once an interview runs past four or five questions, group them into sections — "Daily operations", "When it breaks", "Handover". Choose Add section, name it, and say in a line what that stretch is trying to learn.
They help three ways:
- The builder stays readable when the guide gets long.
- Your colleague sees which part of the interview they are in, so a twelve-question interview does not feel like an interrogation.
- The AI interviewer reads the section description as context for its follow-ups.
Move a question between sections with the section chip on the question card. Deleting a section keeps its questions — they come back out ungrouped. Sections are optional.
Rehearse it before you send it
Preview opens the interview beside the guide and lets you actually take it — speak or type an answer, get the real AI follow-ups, and carry on to the end. It is the fastest way to find out that question three is unanswerable, or that a follow-up instruction is pushing in the wrong direction.
Questions and AI follow-ups alike are read aloud in the same voice your colleagues will hear, with the same Read aloud switch and "Hear it again" button.
Nothing from a rehearsal is saved: no response, no transcript, no recording, and you never appear in your own results. Each question card has its own Preview button, so you can jump straight to the one you are working on.
Tips
- Open-ended beats closed. "Walk me through the last time this broke" gets you knowledge; "Do you check the file?" gets you "yes".
- Allow follow-ups where the knowledge hides. Judgment calls, exceptions and war stories are worth 2–3 follow-ups. Straightforward factual questions need none.
- Keep it to 3–8 questions. A focused interview gets fuller answers than an exhaustive one.
- Ask about people and triggers, not only steps: who they ask, what tells them something is wrong, what they check first.
When the guide is ready, hit Review to check it over and launch it. See Sharing and responses.