Turn a pile of interview transcripts into themes you can defend

Turn interview transcripts into defensible themes with every quote traced back to the person who said it.

For anyone who ran the interviews and now has to say what they mean · 7 steps · 7 min

How it works today.

You export twelve transcripts from Otter or Grain, open a blank document, and start reading. You highlight sentences that feel important and paste them into a scratchpad. By transcript four you have forty quotes and no structure. By transcript eight you are skimming because you think you know the pattern, and the last four interviews get less attention than the first.

You write up five themes from memory, pull a quote for each, and send the deck. Someone asks in the review where theme three came from and you cannot point to more than two sources. You know you heard it more than twice, but you would need another day to prove it.

Before you start.

All of it has to be true, or step one fails in a way that is annoying to debug.

All interview transcripts exported as plain text or docx, with filenames that identify the participant or role.
A generalist agent with file upload (Claude, ChatGPT, or Gemini) and enough context window for all transcripts at once—at least 200k tokens for twelve hour-long conversations.
Access to the original recordings or notes if you need to verify a quote the agent surfaces, because transcription errors do happen.

The steps.

Upload all transcripts in a single conversation
Create a new conversation in your agent and attach every transcript file at once. Name the conversation something you will find again, like the project name and date. Check that the agent confirms how many files it received and that the count matches your folder. If it refuses or truncates, you are over the context limit and need to split into two batches.
02
Ask the agent to extract every distinct point raised by participants
Do not ask for themes yet. You want the raw list of things people said, each one tagged with who said it and which transcript. This step is a structured dump, not analysis. If the agent starts grouping or paraphrasing, stop it and ask again for the verbatim points with citations.
Paste this
```
List every distinct point, complaint, need, or observation raised by participants across all transcripts. For each point, include: the verbatim quote or closest sentence from the transcript, the participant identifier from the filename, and the transcript filename. Do not group or summarise yet. I need the raw list with full traceability.
```
03
Ask the agent to propose themes and show which points belong to each
Now ask it to group the points into themes. Each theme must list the participant quotes that support it, with names. A theme supported by one or two people is not a theme, it is an edge case. If a proposed theme has fewer than four supporting quotes from different people, flag it and decide whether it stays.
Paste this
```
Group the points into themes. For each theme, provide: a one-sentence description of the theme, the list of participant quotes that support it (verbatim, with participant name), and a count of how many distinct participants contributed to this theme. If fewer than four participants mentioned something, call it out as a weak or single-voice theme.
```
Spot-check three quotes against the original transcripts
Pick three quotes from three different themes and search for them in the source files. You are checking that the agent did not rephrase, that the participant name is correct, and that the quote is not taken out of context. If you find a mismatch, ask the agent to regenerate that theme with stricter citation rules.
05
Ask for a summary table of themes with participant coverage
Request a table: one row per theme, columns for theme name, number of supporting participants, and participant names. This is the artifact you will screenshot for the appendix. It makes it obvious which themes have broad support and which are narrow. If any theme shows up with only two names, you probably need to merge it or drop it.
Paste this
```
Create a table with these columns: Theme name, Number of participants who mentioned it, Participant names (comma-separated). Sort by number of participants, descending.
```
06
Export the full theme write-up with inline citations
Ask the agent to write a final document: each theme as a section, a paragraph explaining it, and the supporting quotes underneath as a bulleted list with participant names in brackets. This is the artifact you will edit and send. The agent will try to clean up the quotes for readability—tell it not to.
Paste this
```
Write a final report. For each theme: a heading, a two-to-three sentence explanation of what the theme means, and then a bulleted list of supporting quotes. Each bullet must be verbatim from the transcript, with the participant name in brackets at the end. Do not paraphrase the quotes.
```
Review every theme for coverage and rename weak ones
Read through the final document. Any theme supported by fewer than a third of your participants is either niche or the agent misread the pattern. Decide whether to rename it as an edge case, merge it with another theme, or drop it entirely. The goal is a set of themes you can defend in a room, not a list of everything anyone said once.

What you keep.

Automating the typing does not move the accountability. These stay with a person.

Deciding which themes are strong enough to present and which are edge cases, because the agent counts votes but does not know your research question.
Verifying that quotes are accurate and in context, because transcription errors and agent paraphrasing both happen and you are accountable for what you claim people said.
Writing the implications and recommendations, because the agent can tell you what people said but not what you should do about it.
Choosing which quotes to highlight in the final presentation, because not every supporting quote is equally vivid or persuasive.

Once it works.

The first run is the demo. These are where the time actually comes back.

Run it on a loop

After the first round, you can upload new transcripts to the same conversation and ask the agent to update the themes, so continuous research does not mean starting over every time.

Scale it wider

If you run the same interview protocol across multiple projects, save the theme-extraction prompt as a template and reuse it, so every project gets the same rigor without rethinking the process.

Hand off to another agent

Export the participant-coverage table as a CSV and hand it to someone else doing quantitative analysis, so they can see which themes to weight in a survey or scorecard.

Fire it from an event

Set a calendar reminder one week after the last interview to run this process, so transcripts do not pile up and you do not rely on memory when the research is still fresh.