Use a Claude project to keep the permitted transcripts and coding rules together. Define the categories before asking for a theme summary, then require evidence that lets another researcher check each assignment.
Set up the project around one research question
Open Projects, select + New Project, and give it a name and description. Put the actual task context in the instructions too: Anthropic says Claude does not have access to the project's name and description. Add reference material to project knowledge with +, then use Set project instructions to add the analysis rules. Claude project setup
Choose a question narrow enough to answer from the interviews, such as “Where do first-time users get stuck when booking equipment?” Do not ask for a conclusion about all users when your material describes a few interviews.
Use only material you are allowed to upload. Give each transcript a nonidentifying label, such as Interview-A, and retain section labels or timestamps. Keep the participant identities in your approved research system if the analysis does not need them.
Define a small coding scheme
The following is a fictional starting scheme, not a finding from real interviews:
| Code | Include | Keep separate |
|---|---|---|
| Navigation | Cannot find the booking step | A step that is easy to find but unclear |
| Instructions | Cannot understand what to enter | A system error after valid entry |
| Availability | Cannot find a usable date or item | Uncertainty about the instructions |
Add a category for material that does not fit. Forcing every passage into your starting scheme can hide a problem you did not anticipate.
Project knowledge provides reusable context across the project's chats. Treat important coding rules as explicit instructions or reference material rather than assuming a decision buried in one conversation will always be applied. Project knowledge and instructions
Ask for the evidence matrix before the themes
Analyze the permitted transcripts in this project for this question:
[RESEARCH QUESTION]. Use the code definitions in [CODEBOOK FILE].
First produce an evidence matrix:
transcript label | section/timestamp | short supporting quotation |
assigned code | reason for the assignment | uncertainty.
Use only passages present in the transcripts. Do not invent quotes,
participant details, motivations or prevalence estimates. Use
“Unclassified” when a passage does not fit the codebook.
Then propose themes supported by those rows. For each theme, list
supporting rows and contrary evidence. Distinguish what a participant
said from an interpretation. Mark any recommendation as a proposal.
Check a few rows from each transcript before requesting a longer narrative. Open the quoted passage and read the question that came before it. “That was easy” might refer to finding the item, not completing the booking.
If two passages receive different codes for the same reason, refine the definition and ask for those rows to be reconsidered. Preserve the earlier definition in your research record so you can explain what changed.
Keep contradictions visible
A theme can be useful without being universal. If one participant found the instructions clear while another misunderstood them, keep both observations. Ask what differs in the evidence rather than averaging them into “users had mixed feelings.”
Before sharing, verify quotations, check that transcript labels are correct, and remove unsupported counts. Three mentions from one participant are not three participants. Report the scope of the material you actually analyzed.
Hand off the checked findings
Keep your codebook, source transcripts, and final analysis in the team's approved research record. Send the theme summary with a short explanation of the material covered and the questions that need further interviews.
Optional: read the summary in NOS
If you prepare an HTML review copy, NOS can be a place to read that summary and its evidence references. Keep identifiable interview material in its approved location; an attractive review page does not change who may receive it.