Synthesize interview transcripts and PDFs for research
Last updated: June 10, 2026. Task: Last updated: June 10, 2026 A PM asks “what did users say about checkout?” and the researcher spends the next hour Ctrl+F-ing through seventeen transcripts. That afternoon is common enough that one UX researcher documented it almost word for word before rebuilding her whole process. Claude Code replaces that hour with one typed question, answered with verbatim quotes and file names. This is a guide to Claude Code for researchers who work in words: interview transcripts, PDFs, field notes, and synthesis that has to survive peer review. No code required. I spent years as a PM doing transcript archaeology by hand, and this is the tool I wish someone had handed me. Why Claude Code for Researchers Who Don’t Write Code Claude Code is a chat that runs inside a folder on your computer, which turns out to be exactly the shape of qualitative analysis. Point it at a directory of transcripts and it reads every file, searches across them, and writes new ones: theme summaries, evidence tables, annotated bibliographies. The official docs have a whole section on working in notes and non-code folders, so this is supported use, never a hack. It starts read-only and can only write inside the folder you launched it from, per the security docs. Three things make it better than pasting transcripts into a chat window: Your data stays put. Transcripts live as files in one folder. No uploading ten attachments per conversation. You can re-query. UX researcher Else van der Berg, who runs 50-transcript studies this way, put it plainly: “When my ‘listening for’ changes, I can re-query the data in minutes.” Outputs are files too, which means themes.md and evidence-table.csv exist tomorrow, survive a closed laptop, and can be emailed to a coauthor who has never heard of any of this. One scoping note. If you are an economist who writes R and wants scripted regressions, Paul Goldsmith-Pinkham’s researcher setup series is the better guide. This page is for qualitative and mixed-methods researchers, UX researchers, and grad students whose raw data is language. You have company: 88% of UX researchers already use AI for analysis and synthesis, per an NN/g figure. The open problem is doing it without torching your integrity, which is why half this page is about verification. If you have never touched a terminal, start with the complete guide for non-developers, then come back. Claude Code Interview Analysis: Themes, Quotes, and Contradictions The setup that makes everything else work is boring: one folder, one file per interview, consistent names (p01-maria.md, p02-devon.md). Van der Berg’s insight from her 50-transcript practice is that consistent structure is what makes a folder queryable. Inside a session, type @ to point Claude at any specific file, a habit the course teaches in working with files. 1Set up the folderOne file per interview, consistent names, plus a CLAUDE.md holding your research questions and the citation rule. 2Code in batches of 5 to 8Ask for themes with a verbatim quote, file name, and location behind every claim. 3Hunt contradictionsAsk where participants disagree with each other, and with your own hypothesis. 4Run the verification passClaude searches each quoted string in its source file and flags anything it cannot find. 5Synthesize across batchesMerge batch summaries into themes.md and an evidence table a coauthor can audit. A first-pass prompt that bakes in traceability: Then go looking for trouble on purpose: That second prompt matters because models drift toward agreeable summaries. One researcher described the failure mode precisely: “AI confidently told me the main barrier was pricing, but the real signal was buried in how people talked about trust.” Claude does the mechanical reading. The interpretation, the part your name goes on, stays yours. Andrea Chiarelli, who automated thematic coding for a real evidence review, draws the same line: the machine handles the mechanical part of coding, the researcher keeps the intellectual part. His output format is worth stealing wholesale: a coded evidence table as CSV, a codebook, and a processing log, with “every quote traceable back to its source document.” Ten Transcripts at Once: Parallel Agents for First-Pass Coding Do not dump 50 transcripts into one conversation. The context window fills, quality sags, and you cannot tell where. Practitioners converge on batches of 5 to 8 transcripts per pass, synthesized afterward across batches.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include a local folder structure with consistent naming and a CLAUDE.md file. Data privacy is a major constraint: file contents are sent to Anthropic's servers, requiring de-identification and IRB approval. Model-improvement settings must be checked to prevent training on sensitive data. Verification is mandatory due to high hallucination rates. The tool is not a replacement for CAQDAS software but a first-pass coding aid..
Tools
Published on Agent Work Atlas:
Overview
Last updated: June 10, 2026
Workflow
Last updated: June 10, 2026 A PM asks “what did users say about checkout?” and the researcher spends the next hour Ctrl+F-ing through seventeen transcripts. That afternoon is common enough that one UX researcher documented it almost word for word before rebuilding her whole process. Claude Code replaces that hour with one typed question, answered with verbatim quotes and file names. This is a guide to Claude Code for researchers who work in words: interview transcripts, PDFs, field notes, and synthesis that has to survive peer review. No code required. I spent years as a PM doing transcript archaeology by hand, and this is the tool I wish someone had handed me. Why Claude Code for Researchers Who Don’t Write Code Claude Code is a chat that runs inside a folder on your computer, which turns out to be exactly the shape of qualitative analysis. Point it at a directory of transcripts and it reads every file, searches across them, and writes new ones: theme summaries, evidence tables, annotated bibliographies. The official docs have a whole section on working in notes and non-code folders, so this is supported use, never a hack. It starts read-only and can only write inside the folder you launched it from, per the security docs. Three things make it better than pasting transcripts into a chat window: Your data stays put. Transcripts live as files in one folder. No uploading ten attachments per conversation. You can re-query. UX researcher Else van der Berg, who runs 50-transcript studies this way, put it plainly: “When my ‘listening for’ changes, I can re-query the data in minutes.” Outputs are files too, which means themes.md and evidence-table.csv exist tomorrow, survive a closed laptop, and can be emailed to a coauthor who has never heard of any of this. One scoping note. If you are an economist who writes R and wants scripted regressions, Paul Goldsmith-Pinkham’s researcher setup series is the better guide. This page is for qualitative and mixed-methods researchers, UX researchers, and grad students whose raw data is language. You have company: 88% of UX researchers already use AI for analysis and synthesis, per an NN/g figure. The open problem is doing it without torching your integrity, which is why half this page is about verification. If you have never touched a terminal, start with the complete guide for non-developers, then come back. Claude Code Interview Analysis: Themes, Quotes, and Contradictions The setup that makes everything else work is boring: one folder, one file per interview, consistent names (p01-maria.md, p02-devon.md). Van der Berg’s insight from her 50-transcript practice is that consistent structure is what makes a folder queryable. Inside a session, type @ to point Claude at any specific file, a habit the course teaches in working with files. 1Set up the folderOne file per interview, consistent names, plus a CLAUDE.md holding your research questions and the citation rule. 2Code in batches of 5 to 8Ask for themes with a verbatim quote, file name, and location behind every claim. 3Hunt contradictionsAsk where participants disagree with each other, and with your own hypothesis. 4Run the verification passClaude searches each quoted string in its source file and flags anything it cannot find. 5Synthesize across batchesMerge batch summaries into themes.md and an evidence table a coauthor can audit. A first-pass prompt that bakes in traceability: Then go looking for trouble on purpose: That second prompt matters because models drift toward agreeable summaries. One researcher described the failure mode precisely: “AI confidently told me the main barrier was pricing, but the real signal was buried in how people talked about trust.” Claude does the mechanical reading. The interpretation, the part your name goes on, stays yours. Andrea Chiarelli, who automated thematic coding for a real evidence review, draws the same line: the machine handles the mechanical part of coding, the researcher keeps the intellectual part. His output format is worth stealing wholesale: a coded evidence table as CSV, a codebook, and a processing log, with “every quote traceable back to its source document.” Ten Transcripts at Once: Parallel Agents for First-Pass Coding Do not dump 50 transcripts into one conversation. The context window fills, quality sags, and you cannot tell where. Practitioners converge on batches of 5 to 8 transcripts per pass, synthesized afterward across batches.
Reported result
The source reports that 88% of UX researchers already use AI for analysis. However, it highlights significant integrity risks: citation fabrication rates range from 14.23% to 94.93% when models cite from memory, and 7.7% of AI-supplied 'verbatim' quotes were not found in source transcripts in one audit. No specific success metrics for this exact workflow are provided, only that it replaces manual Ctrl+F searches and compresses coding time.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include a local folder structure with consistent naming and a CLAUDE.md file. Data privacy is a major constraint: file contents are sent to Anthropic's servers, requiring de-identification and IRB approval. Model-improvement settings must be checked to prevent training on sensitive data. Verification is mandatory due to high hallucination rates. The tool is not a replacement for CAQDAS software but a first-pass coding aid.
Try it
Act as a UX researcher using Claude Code to analyze interview transcripts. 1. **Setup**: Create a folder named `interviews` with 5 sample transcripts (e.g., `p01-user.md` to `p05-user.md`). Create a `CLAUDE.md` file containing: "Research Question: What are the main barriers to checkout? Integrity Rule: Only cite files in this folder. Never generate citations from memory." 2. **First Pass Coding**: Ask Claude: "Read all transcripts in this folder. For each theme, provide a summary and 2-3 verbatim quotes with file names and locations. Only quote text that appears word-for-word." 3. **Contradiction Hunt**: Ask: "Where do participants contradict each other about pricing? Show both sides with verbatim quotes. List anything contradicting the hypothesis in CLAUDE.md." 4. **Verification**: Ask: "For every quote in your previous response, search the source file for the exact string. Output a table: Quote, File, Found Verbatim (Yes/No). Flag any not found." 5. **Synthesis**: Ask: "Merge the verified themes into a `themes.md` file and create an `evidence-table.csv` with columns: Theme, Quote, Source File, Location." **Quality Check**: Manually verify 3 random quotes against the raw transcripts. If any are paraphrased or missing, reject the output. **Safeguard**: Do not use real user data. Use dummy transcripts. Confirm no sensitive PII is present before running. If Claude cites a file not in the folder, stop and correct the prompt.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.