Overview
What it is
A long-running Claude agent (Sonnet 4.5) configured with a system prompt, a transcript-loading tool, and a quote-extraction tool. It iterates over a batch of transcripts, clusters themes, pulls verbatim quotes, and flags low-confidence findings before producing a final readout.
When to use it
When you have 10+ transcripts and want a hands-off synthesis run rather than an interactive Claude chat. Best for end-of-cycle research reviews.
Who it’s for
Researchers and PMs comfortable kicking off an agent run and reviewing the output.
Expected output
A markdown readout with: themes (ranked by frequency), supporting quotes (with transcript IDs), confidence per theme, surprises, and recommended follow-ups.
Setup instructions
- Open the Cabana Agents workspace in Claude.
- Duplicate the ‘Research Synthesizer’ agent template.
- Paste the system prompt from research-synthesizer.system.md.
- Connect the transcript-loader tool (points at the Research Drive folder).
- Run with a folder ID as input.
Usage
Good inputs
- • A Drive folder of cleaned transcripts (one per interview).
- • A 2–3 sentence research question.
- • Optional: prior synthesis to compare against.
Example prompts
- Synthesize all transcripts in folder 2026-Q2-onboarding against the question ‘where do new members drop off in week 1?’
Expected outputs
- • A markdown readout (~2 pages) with themes, quotes, confidence, and follow-ups.
Tips & limitations
- • Always review low-confidence themes manually — the agent flags but doesn’t resolve them.
- • Cap batches at ~25 transcripts per run for best coherence.
- • Save the run ID so you can re-open the trace.