qlaude: AI-powered qualitative coding
A research tool where Claude AI proposes, and the researcher decides: turning hours of transcript coding into a rigorous, AI-augmented workflow
Case study
qlaude: AI-powered qualitative coding
The Challenge
Qualitative research is powerful but expensive. The coding phase alone can consume 40-60% of a research project's timeline. For freelancers and small teams, this means either cutting corners (fewer interviews, shallow coding) or blowing budgets. Existing tools like Atlas.ti, NVivo, and Dovetail digitize the process but don't reduce the cognitive load: they're essentially highlighters with databases.
Qualitative coding (the process of reading interview transcripts, tagging meaningful passages with thematic codes, and synthesizing findings) is one of the most time-intensive activities in UX research. A single 60-minute interview can take 3-5 hours to code manually. Multiply that by 10-15 interviews per project, and you're looking at weeks of meticulous, repetitive intellectual labor.
I built qlaude because I was tired of the bottleneck. Not tired of the rigor, tired of the friction. I wanted a tool that understood qualitative methodology well enough to propose codes, suggest highlights, and draft analysis sections, while keeping the researcher firmly in control of every interpretive decision.
"Researchers face a painful choice: spend days on manual coding (accurate but slow), or use generic AI summarization (fast but methodologically shallow). There's nothing in between: no tool that speaks the language of qualitative methodology (themes, codes, codebooks, saturation) while leveraging AI to accelerate the mechanical parts of the work."
The slowness of coding directly limits the scope of research. Teams interview fewer people, code fewer transcripts, and produce thinner insights, not because the method isn't valuable, but because the tooling doesn't respect it.
Researchers copy-paste transcripts into spreadsheets. They manually read every line, highlight passages, assign codes, then re-read to check consistency. Theme generation happens on sticky notes or whiteboards. Report writing means re-reading all coded passages again to synthesize. Each phase restarts the same reading process from scratch.
Research
Practice-Based Needs Analysis
The Reading Bottleneck
60% of coding time is spent re-reading: first to understand, then to code, then to verify, then to synthesize. AI could handle the 'propose' step, letting researchers focus on 'decide.'
Methodology Matters
Generic AI summarization produces outputs that look like research but aren't: no codebook structure, no traceability to source text, no inter-coder consistency. Researchers need AI that speaks their language.
3-5 hours per transcript
Average manual coding time for a 60-minute interview. With 10+ interviews per project, this single phase can consume 30-50 hours.
Tool Landscape Analysis
AI is Bolted On, Not Designed In
Tools adding AI features (Dovetail's summaries, NVivo's auto-coding) treat it as a feature, not a workflow. The AI doesn't understand codebook structure, doesn't reference specific text, doesn't explain its reasoning.
No Tool Bridges Coding and Analysis
Coding tools stop at the codebook. Report writing happens elsewhere. The researcher must mentally reconnect coded passages to narrative analysis: exactly the kind of synthesis AI does well.
Traceability is Non-Negotiable
Academic and professional research requires that every claim trace back to source data. Any AI assistance must preserve this chain: quote → code → theme → insight.
Key Insights
AI should propose, researchers should decide
The fundamental interaction model. Claude suggests themes, highlights passages, drafts analysis, but every output is a proposal that the researcher reviews, edits, accepts, or rejects. The human always has the final word.
Methodology is the interface
The tool must speak qualitative research language natively: themes (KDs), codes, codebooks, highlights, saturation. Not 'tags' and 'summaries' but actual research constructs that map to established methodology.
Every AI output must be traceable
When Claude proposes a highlight, it must reference the exact text, the exact segment, and explain its reasoning. When it drafts analysis, it must cite specific participant quotes. No black boxes.
The codebook is the contract
AI suggestions are constrained by the codebook the researcher defines. Claude doesn't invent categories: it works within the framework the researcher establishes, ensuring consistency across transcripts.
Speed without shortcuts
The goal isn't to skip qualitative coding, it's to remove the mechanical friction so researchers can focus on interpretation. Every hour saved should increase depth, not reduce rigor.
Goals & Principles
AI-assisted codebook generation
Given a research scope, Claude proposes an initial set of themes and codes: a structured starting point the researcher refines, not a finished product.
Intelligent transcript coding
Claude reads transcripts against the codebook and proposes highlights: exact text passages matched to specific codes, with reasoning for each suggestion.
Automated analysis drafting
From coded data, Claude generates publishable-quality analysis sections: flowing prose with participant quotes, pattern identification, and prevalence quantification.
Full report synthesis
Aggregate all coded data across interviews into a complete thematic analysis report, structured by the codebook, with cross-participant pattern analysis.
Guiding Principles
Researcher sovereignty
Every AI suggestion is reviewable and reversible. The researcher's judgment overrides AI proposals at every stage.
Methodological transparency
AI reasoning is always visible. When Claude proposes a code, it explains why. When it drafts analysis, it cites sources.
Progressive assistance
AI helps more as the project matures. Early stages (codebook design) get light suggestions. Later stages (report writing) get heavier drafting, because the researcher has established the analytical framework.
Solution Design
AI-Powered Codebook Scaffolding
Starting a codebook from scratch is daunting. Researchers stare at a blank canvas, knowing they need themes and codes but unsure where to begin until they've read transcripts.
Claude reads the research scope and streams a structured codebook proposal: 4-6 themes with 2-4 codes each, following KD notation (KD1.1, KD1.2). The researcher reviews and restructures using a drag-and-drop codebook manager built with React Flow.
This isn't auto-coding, it's scaffolding. The AI provides a starting hypothesis that the researcher refines through engagement with the data. It reduces the 'blank page problem' while keeping the researcher in analytical control.
Transcript Highlight Proposals
Reading transcripts line by line to find codable passages is the biggest time sink. The researcher knows what they're looking for (the codebook defines it), but finding it requires reading everything.
Claude receives the full transcript with segment IDs, the complete codebook with UUID mappings, and any existing highlights (to avoid duplicates). It proposes 5-15 highlights per transcript, each with: exact verbatim text, the segment it belongs to, the code it matches, and 1-2 sentences of reasoning.
The prompt engineering here was critical. Claude must quote text exactly (no paraphrasing), reference specific segment UUIDs (for accurate positioning), and respect existing highlights (no duplicates). Structured output via Zod schema ensures consistent, parseable responses streamed in real-time.
Analysis Generation: From Codes to Prose
After coding, researchers face another manual phase: reading all passages tagged with a code and writing narrative analysis. This requires re-engaging with every quote, identifying patterns, and synthesizing findings into prose.
Two levels of analysis generation. Per-code analysis: Claude writes flowing prose about all quotes tagged with a specific code, identifying patterns, quantifying prevalence, and weaving in verbatim quotes with participant attribution. Full report: Claude synthesizes the entire project into a publishable thematic analysis with cross-theme connections.
The prompt constraints are where the design work lives. Claude is instructed to: write in flowing prose (no bullets), quantify prevalence ('three of five participants...'), always cite verbatim quotes with attribution, surface minority views and contradictions, and write in the same language as the participants. These constraints produce output that reads like rigorous research, not AI summary.
Outcomes & Impact
6 distinct AI interaction points across the full research workflow
Codebook scaffolding in under 30 seconds (vs. 1-2 hours manual)
5-15 highlight proposals per transcript, streamed in real-time
Full thematic analysis report generated from coded data in minutes
Structured output via Zod schemas ensures 100% parseable AI responses
Researchers spend time on interpretation, not transcription scanning
AI proposals with reasoning create a dialogue between researcher and tool
Codebook consistency across transcripts is maintained by the AI working within researcher-defined codes
Report drafts include proper citations and prevalence quantification out of the box
The tool respects qualitative methodology (themes, codes, saturation) as first-class concepts
Reflections
What Worked
Constraint-Driven Prompt Engineering
The strict writing guidelines in analysis prompts (no bullets, cite quotes, quantify prevalence, surface minorities) produced dramatically better output than open-ended generation. Telling the AI how to think like a qualitative researcher made it produce output that actually reads like qualitative research.
Structured Output Eliminated Parsing Fragility
Using Zod schemas with Vercel AI SDK's streamObject meant every AI response was typed and validated. No regex parsing, no hoping the AI formatted correctly. This was essential for features like highlight proposals where the response must contain exact UUIDs and segment references.
The 'Propose, Don't Decide' Model Built Trust
By framing every AI interaction as a suggestion rather than an action, researchers could engage critically with AI output rather than feeling replaced by it. The tool earns trust through transparency: every proposal shows its reasoning.
Challenges
Prompt Engineering is UX Design
The hardest design decisions weren't in the interface, they were in the prompts. Getting Claude to quote text exactly (not paraphrase), reference specific segments (not approximate), and write in the participant's language required iterative prompt refinement that felt more like interaction design than engineering.
Balancing Assistance with Agency
Too much AI and researchers feel passive. Too little and the tool doesn't save time. Finding the right level of suggestion (enough to accelerate, not enough to override) required constant calibration of how much the AI proposes at each stage.
Key Learnings
AI augmentation works best when it respects domain methodology: not replacing expertise but amplifying it within established frameworks
Prompt engineering is a design discipline: the constraints you give AI shape its output as much as any interface decision
Structured output (schemas, typed responses) is non-negotiable for production AI tools: 'hope the AI formats it right' doesn't scale
The best AI tools make the human more capable, not less necessary: qlaude makes researchers faster while keeping them essential
Building your own tools teaches you more about AI's capabilities and limits than any amount of reading: the gap between 'AI can do this' and 'AI can do this reliably' is where the real design work lives