Blog · · 7 min read
Turn interview notes into findings you can check
#The situation
You ran six customer interviews over a week. Each one left a page of notes: who they are, what they said about setup, about price, why they left or why they stayed. A few lines are in quotation marks, because you wrote them down word for word.
Now someone asks the obvious question: so why are people leaving? You know the answer is in those six pages. You paste them into an AI assistant, ask for a summary, and get back five tidy bullets. They sound right. But one of them says "most users found the price too high", and you only remember one person saying that. Was it most? Which one said it? You go back to the notes to check, and the summary saved you nothing.
A summary you can't check is an opinion. What you want is a findings report where every claim carries its count and every quote carries its source.
#Ask the AI
Paste this, then your interview notes underneath, each with its filename on the line above it. It works in ChatGPT, Claude, Gemini, or an agent that can read the folder directly.
Below are my notes from customer interviews, one file each, each starting
with its filename. They live in an interviews/ folder. Write a findings
report I'll save next to that folder.
What I wanted to learn: <your question, in one line>
Name it: <today's date YYYY-MM-DD>-findings.md
The file:
- Frontmatter: title, date, interviews (how many), question (the line above).
- A # heading, then "## Answer": three or four sentences that answer the
question directly.
- If one finding should change a decision now, put it in a > [!important]
callout right under the Answer.
- "## At a glance": a table with columns Finding, Interviews, Strength.
Finding links to its section below, like [Setup is too slow](#setup-is-too-slow).
Interviews is a count like "4 of 6". Strength is strong (half or more),
some (two or more) or single (one).
- One "## <finding>" section per finding, strongest first: two sentences on
what we heard, then up to three quotes as > blockquotes. After each quote,
link to the interview it came from, at the heading where it was said:
[P3, Setup](interviews/<filename>#setup).
- "## Where they disagreed": who said opposite things, and what might explain it.
- "## What we didn't ask": - [ ] questions that could change the answer.
Rules:
- Quotes are word for word from the notes. Never paraphrase inside a quote.
- Count an interview for a finding only if its notes show it.
- Anchors: the heading in lowercase, spaces as hyphens, punctuation removed.
Keep finding headings short and without punctuation so the links work.
- No advice beyond the Answer and the callout.
Reply in the chat with the filename only. Then the file, in one code block
fenced with four backticks.
It works best if your notes share the same few headings (Background,
Setup, Pricing, Leaving, Wishes). That's what makes a link like
#setup land in the right place in every interview.
Four lines do most of the work.
"Interviews is a count like 4 of 6." Many users and some users hide the number. 4 of 6 doesn't, and it also tells you when a finding rests on one person.
"Word for word. Never paraphrase inside a quote." Asked to summarise, an assistant smooths. A smoothed quote sounds stronger than what was said, and it's the line that ends up on a slide.
"Link to the interview, at the heading where it was said." This turns every quote into something you can check in one click instead of trusting.
"What we didn't ask." Six interviews never answer everything. Asking for the gaps keeps the report honest about what it can't tell you, and gives you the questions for the next round.
#What comes back
We tried it with six made-up but realistic interviews for a recipe-costing
app for small bakeries, called Crumb. The question: why do bakeries stop
during or after the trial, and what makes the ones who stay, stay? The
assistant replied with the filename, 2026-10-14-findings.md, and this
file. It's shortened for the page: one disagreement and two open questions
are cut.
---
title: Why bakeries stop using Crumb, and why some stay
date: 2026-10-14
interviews: 6
question: Why do bakeries that try Crumb stop during or after the trial, and what makes the ones who stay, stay?
---
# Why bakeries stop using Crumb, and why some stay
## Answer
Most of the bakeries that stopped (P1, P3, P5) ran into setup first: typing
recipes by hand, entering everything in grams, or an import that failed. P6
got through setup in an hour and left because Crumb could not hold packaging
and delivery costs. Price was not the objection for four of the six. The two
who stayed (P2, P4) got through setup and then used Crumb to catch cost
changes: P2 checks his margin report every week, and P4 found two best
sellers selling under cost.
> [!important]
> Price is not what is losing these bakeries. Four of six had no objection to
> it, and P5 said she would pay three times as much. Setup is where most
> trials end, so a decision to change the price should wait.
## At a glance
| Finding | Interviews | Strength |
|---|---|---|
| [Price is not the reason](#price-is-not-the-reason) | 4 of 6 | strong |
| [Setup is too slow](#setup-is-too-slow) | 4 of 6 | strong |
| [Seeing margin changes keeps people](#seeing-margin-changes-keeps-people) | 2 of 6 | some |
| [Needs Crumb does not cover](#needs-crumb-does-not-cover) | 2 of 6 | some |
| [They want a short list](#they-want-a-short-list) | 2 of 6 | some |
| [Price updates went unfound](#price-updates-went-unfound) | 1 of 6 | single |
## Price is not the reason
P2, P4, P5 and P6 had no problem with the price, and two of them left anyway.
P1 also called $19 a month fine, but only "if it actually saved me an evening
a week".
> "Nineteen dollars is nothing. One wrong quote to a restaurant costs me more than a year of it."
[P2, Pricing](interviews/2026-10-06-p2-tomas.md#pricing)
> "If it worked for both shops I'd pay three times that."
[P5, Pricing](interviews/2026-10-09-p5-marta.md#pricing)
> "I'd pay. That wasn't it."
[P6, Pricing](interviews/2026-10-10-p6-dev.md#pricing)
## Setup is too slow
P1, P3 and P5 stopped at setup: hand-typing recipes, a grams-only entry
screen, and an import that failed on merged cells. P4 also found setup slow
(about two weeks of evenings) but thought it was worth it, and P2 got past
setup by importing 60 recipes from his spreadsheet in an afternoon.
> "I spent a Sunday typing in ingredients and I got through nine recipes. Nine. I have forty."
[P1, Setup](interviews/2026-10-05-p1-rosa.md#setup)
> "It asked me for grams of everything. I bake in cups. I gave up on the second recipe."
[P3, Setup](interviews/2026-10-07-p3-ana.md#setup)
> "The import was the only reason I stayed past the trial."
[P2, Setup](interviews/2026-10-06-p2-tomas.md#setup)
## Seeing margin changes keeps people
Both active customers use Crumb when ingredient prices move. P2 updates flour
and butter prices when the invoice comes in and checks the margin report
every week, and P4 found two best sellers under cost after butter went up.
> "That was the moment. I raised them the same day."
[P4, Use](interviews/2026-10-08-p4-kenji.md#use)
## Needs Crumb does not cover
P5 and P6 left because Crumb did not match a need they had. P6 could not add
packaging and delivery cost per order, and P5 wanted two locations with
different prices for the same product. P5's manager also already had a
spreadsheet that worked.
> "The recipe cost is the easy part. The box and the courier are where I lose money."
[P6, Leaving](interviews/2026-10-10-p6-dev.md#leaving)
> "It's one person's job and that person already has a spreadsheet that works."
[P5, Leaving](interviews/2026-10-09-p5-marta.md#leaving)
## They want a short list
P1 and P3, both of whom left before finishing their recipes, asked for a
plain answer instead of a full costing tool.
> "Just tell me which three things I'm losing money on. I don't need the rest."
[P1, Wishes](interviews/2026-10-05-p1-rosa.md#wishes)
> "Honestly, a list: this costs you this much, sell it for at least this."
[P3, Wishes](interviews/2026-10-07-p3-ana.md#wishes)
## Price updates went unfound
P1 stopped after three weeks believing she had to change a butter price in
every recipe. Crumb updates every recipe when an ingredient price changes,
but she did not find that.
> "The supplier changed the butter price and I had to go change it in every recipe. Or I thought I had to."
[P1, Leaving](interviews/2026-10-05-p1-rosa.md#leaving)
## Where they disagreed
- **Price.** P3 said "Nineteen a month is more than my stall fee at one of
the markets" and would pay about $40 once. P2, P4, P5 and P6 had no
objection. P3 is a home baker with 12 products. P2 is wholesale on thin
margins where one wrong quote costs more than a year of Crumb.
- **Setup.** P2 finished 60 recipes in an afternoon and P4 took two weeks,
yet both stayed. P1 and P3 gave up after 9 and 1 recipes. P2 already had a
spreadsheet to import. P4 had sub-recipes, which he called "the whole
point". This could be about whether a bakery has something to import, or
about how soon it sees a payoff.
## What we didn't ask
- [ ] Did P1, P3 or P6 ever see the margin report, or any result from their data?
- [ ] Would P3 have continued if Crumb accepted cups, or if "yield" were explained?
- [ ] Did P5's manager try the import again, and would a fix for merged cells have changed the decision?
- [ ] How close did P2 and P4 come to leaving during their own trials?
This is a good report, and the best thing in it is the callout. You'd probably have walked into the interviews thinking price was the problem. The report says it isn't (four of six said so, one offered to pay three times as much) and that setup is where trials end. That's a decision you might otherwise have got wrong, backed by quotes you can open.
It also kept a single-person finding at 1 of 6 instead of dropping it. Price updates went unfound rests on one interview, but it's the cheapest fix on the list: the feature exists, and P1 never found it.
Two things are worth a second look.
Check the counts against the notes, not against your memory. The report puts P4 under Setup is too slow, although P4 stayed. That's right (the notes say "took about two weeks of evenings") but it's the kind of call you should see before you quote 4 of 6 to anyone.
"Gave up after 9 and 1 recipes" is one step too far. P3's notes say she gave up on the second recipe and never finished one. Small, and exactly the sort of thing the quote links let you catch: one click to P3, Setup and you can see it.
#Save it as a file
Save it as 2026-10-14-findings.md next to the interviews folder, in one
crumb-research folder, and connect that folder in MD Flow (what browsing
a folder of Markdown looks like).
#Why that matters next week
Every claim is a click from its evidence. The report reads as a document: the frontmatter as a header, the callout you can't miss, the at-a-glance table. A link in the table takes you to that finding's section. A quote's link opens the interview note right in MD Flow, at the heading where it was said, so "is that really what Ana said?" takes one click instead of a search through six files.
The notes and the report stay together. Next week someone asks about the import. The notes are in the same folder as the report, not in a chat you'd have to scroll back through. Search by filename is free in MD Flow; with Pro, search also reaches inside the files, so spreadsheet finds every interview that mentions one, each with the line and its number.
The next round has a starting point. The open questions at the bottom
are the first draft of the next interview guide. After round two, save the
new notes in the same interviews folder and run the prompt again. The new
report keeps the same table, and you can open last month's report next to
it and see which counts moved.
It's the same file in any assistant. Paste the notes into ChatGPT today and Claude next month; the format holds. If you use an agent that can read your folders, point it at the folder instead of pasting.
An interview is where you hear it. The findings file is what you act on, and the folder of notes behind it is why anyone should believe it.