Thematic analysis

Thematic analysis is a way of finding, analysing and reporting patterns of meaning, called themes, across qualitative data: interviews, focus groups, open survey answers, diaries. Virginia Braun and Victoria Clarke's 2006 paper "Using thematic analysis in psychology" set out the version most researchers now follow, and it's used well beyond psychology, in health, education, business and the social sciences.

A theme isn't a topic. "Experiences of night shifts" is a topic: it collects everything said about night shifts. "Tiredness as proof of loyalty" is a theme: a pattern of shared meaning built around one central idea, which tells the reader something about the data. Stopping at topics is a common weakness, and one Braun and Clarke warn against.

Braun and Clarke describe six phases. They're recursive, not a checklist: you'll go back from reviewing themes to coding, and from writing up to naming.

• 1. Familiarisation. Read and re-read the data, listen to the recordings, and note first ideas. Transcribing is part of it: checking an automatic transcript line by line against the recording does much the same work.
• 2. Coding. Work through the whole dataset and label each passage that matters to your research question with a short code. Codes can be semantic, close to what was said, or latent, about the assumptions under it.
• 3. Generating initial themes. Cluster codes that share a central idea into candidate themes, and gather the data for each.
• 4. Reviewing themes. Check each theme against its coded extracts, then against the whole dataset. Merge, split or drop the ones that don't hold.
• 5. Defining and naming themes. Say in a few sentences what each theme is about and where it stops, and give it a name that carries its idea.
• 6. Writing up. Tell the analytic story with vivid extracts, and relate it to your research question and the literature.

Braun and Clarke distinguish three broad kinds. Knowing which one you're doing settles most questions of method, software included.

• Reflexive thematic analysis, their own approach. Your subjectivity is a resource, coding is organic and changes as you go, and themes are generated by you, not found in the data. There's no codebook to agree on and no inter-rater reliability.
• Codebook thematic analysis, such as framework analysis and template analysis. A structured codebook, often set up early, common in applied and team research with set questions to answer.
• Coding reliability thematic analysis. A fixed codebook applied by several coders, with their agreement measured (Cohen's kappa, for example) as a sign of quality.

In late 2025 Braun and Clarke, with Tanisha Jowsey, Deborah Lupton, Michelle Fine and more than 400 qualitative researchers, published an open letter in Qualitative Inquiry: "We reject the use of generative artificial intelligence for reflexive qualitative research." Their case is that reflexive analysis depends on a positioned researcher making meaning, which a language model doesn't do, and that generative AI raises social and environmental concerns besides. If you're doing reflexive thematic analysis, take it seriously: your supervisor, examiners or reviewers may well agree.

AI is less contested in transcription, in codebook and coding reliability work, in rapid applied projects, and as a check after you've coded: a second reader whose codes you set against your own. Whatever you use, describe it in your methods section and check your institution's policy.

Analyse in Claude, under a finished transcript, is built for that second-reader use. It copies the transcript under a prompt that asks for codes, candidate themes and a codebook, with every quote word for word and timestamped so you can check it against the recording, and it presents them as candidates for you to review, not findings. Nothing is sent until you paste it into Claude yourself.

This is the prompt Analyse in Claude uses. Copy it into a new Claude chat, put your research question in place of the second paragraph, and paste your transcript between the tags. Replace real names first: the transcript goes to Anthropic when you send it.

Please do a first pass of thematic analysis on the interview transcript below: initial codes, then candidate themes. I'll review and develop them myself.

I haven't given a research question, so code inductively, from what the participants say.

1. Initial codes. Work through the transcript in order. For each code: a short name, a one-line definition, and the passages it applies to, by timestamp and speaker.
2. Candidate themes. Group the codes into 3 to 6 themes. For each: a name, its central idea in one or two sentences, the codes it brings together, and two or three quotes, with timestamp and speaker.
3. A codebook, as a table: Theme | Code | Definition | Example quote | Timestamp.
4. What doesn't fit: passages that contradict a theme, or that no code covers.

Ground rules:
- Quote word for word, with the timestamp and the speaker, so I can check each quote against the recording. Never put a paraphrase in quotation marks.
- The transcript was made automatically. A word that makes no sense may be a transcription error: point it out instead of building on it.
- The interviewer's questions are context, not data. If the speakers aren't named, say which one you take to be the interviewer.
- These are candidate codes and themes for me to review and revise, not findings.

<transcript>
(Paste your transcript here.)
</transcript>

Recordings hold names, health, work and family. This page's transcriber runs in your browser: the recording and the transcript stay on your computer, and it works with Wi-Fi off. The data-flow statement is written to be quoted in an ethics application or a DPIA.

Sending a transcript to any AI service, Claude included, is a separate step your ethics approval and consent forms have to cover. Rename the speakers (Interviewer, P1) and correct names in the text before you do.