The Agent extracts every meaningful statement, clusters them by sentence embedding in the sandbox, and proposes 5–7 themes for you to confirm. Each theme card lists frequency, distribution across participants, and 3–5 verbatim quotes. Save the framework — your follow-up study auto-codes against the same taxonomy.
Each statement gets embedded, clustered inside the sandbox, and the Agent proposes themes for you to confirm. Each point is a participant statement; its color is the cluster it landed in.
Not "AI summary themes". Each theme card cites 3–5 verbatim quotes from different participants — the lead researcher audits, the analysis is auditable.
“I didn't realize I could just paste a folder and have it written about until day five.”
“First login I just stared. There were nine panels open and no one telling me where to start.”
“I need a per-task cost number before I can let anyone on my team use this with our data.”
Themes + definitions + coding rules are written to a codebook.json in your Drive's research folder. Q2 starts by reading the json, then codes the new batch against the same themes and produces a delta report.
{
"themes": [
{ "id": "T1", "label": "Onboarding overwhelm", "n": 28 },
{ "id": "T2", "label": "Wants keyboard control", "n": 19 },
{ "id": "T3", "label": "Aha late", "n": 24 },
{ "id": "T4", "label": "Pricing anxiety", "n": 11 },
{ "id": "T5", "label": "Audit-trail trust", "n": 14 }
]
}Text, Markdown, DOCX, or PDFs work. If you have audio, run /tools/meeting-minutes first to transcribe — that's the cheaper / faster path. The Agent expects each transcript to identify the participant in the file name or a header line.
Let the themes emerge from the data (inductive), code against your existing framework (deductive), or compare to last quarter's framework (longitudinal). For inductive, the Agent shows proposed themes in chat before rendering — you rename, merge, or split. Nothing is finalized without your sign-off.
The PDF lands in your Drive: theme cards, a 2D map showing how themes relate, a quote-anchor wall (3–5 verbatim excerpts per theme), and an appendix listing every coded statement. The Agent also writes a `codebook.json` to the same Drive folder — next quarter's round reads it and codes against the same themes.
Cheap qualitative tools spit out theme names without grounding. The Agent's theme cards each carry 3–5 verbatim excerpts from different participants — you can read why the theme exists, not just trust a label. The codebook appendix lists every coded statement so the lead researcher can audit every assignment.
Sentence-embedding clustering happens inside your workspace-scoped sandbox with an open embedding model. Transcripts never leave the sandbox boundary; only the named outputs (PDF, framework JSON) land in your Drive. We do not train on customer interview content.
When you reject a proposed theme in chat ("merge these two", "this label is wrong"), the decision is logged in the framework JSON. If a future round tries to reintroduce a similar cluster, the Agent surfaces the prior rejection as context — you can confirm the new round changes things or apply the same rejection consistently.
Most qualitative tools start at zero every project. Here the codebook persists in your workspace. Quarter 2 picks up Quarter 1's themes, adds what's new, retires what's gone — and the Agent shows you the diff. Over a year you have a research framework that captures the evolution of how users talk about your product, not 4 disconnected studies.
Tell it in chat. Common edits: "merge themes A and B — they're the same finding from two angles", "split theme C — there are two distinct sub-themes hiding here", "rename theme D to 'X' — that's how this audience talks about it". The Agent re-renders the PDF with your edits and logs the decision in the framework. The Agent's job is to surface candidate structure — yours is to decide what the data is actually saying.
Sign in, hand it over to the Agent — the finished file lands in your Drive.