A knowledge interview is an AI-moderated conversation with an experienced colleague, shared as a single link, that asks follow-up questions and turns the answers into reusable operational knowledge.
It exists because the alternative does not work. Asking an experienced colleague to write down what they know turns them into an author, which is neither their job nor usually their talent, and the document that comes back is a sanitised version of the procedure rather than the judgment underneath it. A conversation can ask for the last real example, probe the exception, and keep the reason.
Updated · Nuggetz
What makes it an interview rather than a set of questions?
The follow-up. A fixed set of questions can only collect what its author already thought to ask.
When a colleague says an escalation is "usually" fine, the useful knowledge is in the word usually — and only a second question gets it. Nuggetz asks that question: it follows up when an answer contains something worth exploring, and redirects when the conversation drifts away from what you set out to learn. Everything else about the format is chosen to make that affordable. It is asynchronous, so nobody books a meeting with the busiest person on the team. It runs from one link with no account, so participation costs a click. Answers can be spoken, because most people say more in two minutes than they will type in ten.
How does it compare with forms, recorders and wikis?
Each of these is better than Nuggetz at something, and the honest comparison is about which part of the job each one is built for. Every competitor stops somewhere along the same chain: interview, transcript, nuggets, corroboration and conflicts, handbook.
How forms, screen recorders, wikis and Nuggetz each handle undocumented operational knowledge
Forms
Screen recorders
Wikis and SOP tools
Nuggetz
Captures judgment and exceptions
Only what the question anticipated
Only what happened on screen
Only if the author writes it down
Asks for the last real example, then follows up
Needs the expert to author
No, but they self-edit
No
Yes — this is the bottleneck
No, they answer questions
Corroborates across people
Counts answers
No
No
Compares nuggets across everyone who answered
Shows conflicts and gaps
Shows spread, not disagreement
No
Last edit wins
Named and left visible in the handbook
Output linked to evidence
Raw responses
The recording
No
Every instruction links to the answer behind it
Best when
You already know the questions and want counts
The work is click-by-click on one screen
The knowledge is already written and needs a home
The work depends on judgment nobody has written down
Screen recorders are the right tool for click-by-click procedures with screenshots, and we would not try to beat them at it. Wikis are the right home for knowledge that is already written — which is why a Nuggetz handbook exports as Markdown into the one your team already keeps, rather than asking them to read it here.
What do you actually get at the end?
Nuggets, and a handbook built from the ones that settled.
Each answer is cut into nuggets — self-contained pieces of operational knowledge, each typed as a how-to, a judgment call, a gotcha, a why, or a who and where. Cutting a conversation up this way is what makes two people's answers comparable at all. Once more than one colleague has answered, every nugget carries a state: confirmed, single source, conflict or gap. Confirmed means several people independently said the same thing — internal corroboration, never a check against the outside world. The settled nuggets become a handbook, and the conflicts and gaps stay visible in it under "Still unresolved" rather than being flattened into a confident sentence. Every term here is defined on the glossary, and the five steps that produce them are on how it works.
What does one actually look like, end to end?
One question, asked of someone who has run every month-end close for fourteen years. Watch the interview push back when the answer wanders, then follow up when it lands on something real.
Question
What goes wrong on the close, and how do you know it's about to?
Answer
Well, when we migrated the ERP two years ago, that was a nightmare from start to finish, the mapping was wrong on about forty accounts and we…
Redirect
That's useful background. Coming back to a normal close though — what goes wrong, and how do you know it's coming?
Answer
The FX rates. The rate table loads at six in the morning. If you run the revaluation before it's loaded — and on day one you're in early, that's rather the point — it uses yesterday's rates. And it doesn't tell you. No error, no warning, it just runs. Then everything's out by a small amount, and it's small enough that you don't spot it, and you lose a day finding it.
Follow-up
So how do you know it's happened?
Answer
I check the rate date on the revaluation report before I look at anything else. Ten seconds. If it shows yesterday's date, stop and wait.
What the interview took from that
Gotcha · Confirmed
The FX rate table loads at 06:00 and the close uses stale rates before it
A revaluation run before 06:00 uses the previous day's rates. There is no error and no warning — it simply runs. The resulting differences are small enough not to be noticed and take about a day to trace. Day one is exactly when people are in early.
“No error, no warning, it just runs.”
“I lost most of a day to this in my second month. Nobody had told me and there was nothing to read.”
Confirmed, because a second colleague independently described the same trap — not because a model judged it likely. Two people, two separate interviews, and the quotes stay attached so anyone can check.
And what the handbook says
Checks that save a day
Before looking at anything else on day one, check the rate date on the revaluation report. It takes ten seconds.
If it shows the previous day's date, stop and wait for the rate table to load before running the revaluation.
Ten seconds of checking, against a day of tracing differences that were never wrong — only computed at yesterday's rates. Nobody had written it down, because it had long since stopped surprising the person who knew it.
The part a document written by one person cannot do
In the same set of interviews, two colleagues who had closed the books together for years turned out to be applying different thresholds — each certain it was the policy, neither ever asked.
Controller, fourteen years
£500
“Anything over five hundred on an account, I chase it. I think I got it from my predecessor.”
Management accountant
£2,500
“Chasing a six-hundred-pound difference costs more in my time than the difference is worth.”
That is a conflict, and it stays visible under “Still unresolved” rather than being flattened into one confident sentence. Only confirmed and single-source knowledge is allowed to instruct anyone; a conflict is by definition not settled enough to tell a third person to do.
Frequently asked questions
What is a knowledge interview?
A knowledge interview is an AI-moderated conversation with an experienced colleague, shared as a single link, that asks follow-up questions and turns the answers into reusable operational knowledge. It is asynchronous: the colleague answers whenever it suits them, by speaking or typing, and no meeting is booked.
How is a knowledge interview different from a form?
A form can only collect what its author thought to ask. A knowledge interview asks a follow-up when an answer contains something worth exploring — when someone says an escalation is "usually" fine, it asks what makes it unusual. That second question is where the reusable knowledge is.
Who runs the interview — a person or the AI?
You design the guide and the AI moderates. It asks one question at a time, follows up, and redirects when the conversation drifts. Nobody has to sit in the room, which is why it scales past the one expert whose calendar you could actually book.
How is this different from recording a meeting?
A recording produces a transcript of one conversation. A knowledge interview produces comparable claims from several people, so it can tell you where two experienced colleagues disagree — which is a finding a transcript cannot contain.
Do I need to interview more than one person?
No. One experienced colleague is enough to begin. Interviewing more is what turns a single account into knowledge marked confirmed, and what surfaces conflicts between how two people actually do the work.
What comes out of a knowledge interview?
Nuggets — self-contained pieces of operational knowledge, each typed as a how-to, judgment call, gotcha, why, or who and where — and, from the settled ones, a handbook whose every instruction links back to the answer behind it.
Try it on something nobody has written down
Pick one thing your team depends on and has never documented — an escalation procedure, a handover, the way a high performer decides — and run an interview about it.
