Claude Code segments 50 interview transcripts
Van der Berg reports converting more than 50 structured interview files into a classified CSV and filterable HTML dashboard, while retaining her own opportunity mapping and interpretation.
Original prompt
Shared by the source author. See source notes for context.
Weight my manual notes higher than the AI transcript summary when identifying insights and opportunities.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Else van der Berg
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Van der Berg reports converting more than 50 structured interview files into a classified CSV and filterable HTML dashboard, while retaining her own opportunity mapping and interpretation.
- Human involvement
- Van der Berg takes manual notes, defines categories and retains opportunity-tree judgment after the generated tree fails. Classification accuracy is not independently measured.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Store one interview file with profiling data, manual notes and transcript.
- Define the current customer profile and human-led classification categories.
- Extract and classify interview attributes into a CSV.
Else van der Berg organises more than fifty customer interviews as one Markdown file per person, with profiling data, manual notes and the full transcript. Company files define the current ideal customer profile, goals and product context. Claude Code reads this workspace when the research focus changes.
She gives her own notes greater weight than automatic transcript summaries and defines analysis categories from patterns she has observed. Claude Code extracts profile attributes and classifies friction, frequency, severity and desired improvements into a CSV. It then builds a Chart.js HTML dashboard with ordered charts, segmentation filters and a table retaining free-text details.
Van der Berg uses the visualisations to challenge an existing opportunity solution tree. When she asks the agent to draw that tree directly, the result has poor framing and arbitrary parent-child relationships, so she keeps that judgment herself. The completed CSV and dashboard therefore support research interpretation rather than replacing it. The useful structure is to preserve raw transcripts beside human notes and categorical outputs, allowing a changed customer profile to trigger another query. The account does not provide an independent coding-agreement measure or prove that generated classifications are bias-free.
Prompt
Original prompt
Source: Blog · Else van der Berg
Frequently asked questions
Why retain the full transcript?
It lets the researcher return to what was actually said when the profile or research question changes.
How are automatic summaries weighted?
The author instructs Claude to give her manual notes more weight than AI transcript summaries when identifying insights.
Did it replace the opportunity tree?
No. The generated tree had poor framing and arbitrary relationships, so she retains human opportunity mapping.
More details
Remaining steps
- Build a filterable HTML dashboard while retaining free-text details.
- Use the evidence to review the opportunity tree rather than delegate its framing.
About this account
Claude Code reads a structured interview corpus, extracts and classifies data, then generates a CSV and interactive dashboard for repeated segmentation.
Firsthand account by Else van der Berg; original publication date verified against Primary Substack date Feb 04, 2026. Outcomes are the author’s reports, not independently reproduced. Original CLAUDE.md instruction shared by the author; not the full extraction or dashboard prompt.
Credit: Else van der Berg
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · Windsurf · Markdown interviews · CSV · Chart.js
Interview synthesis · Customer segmentation · Human-led categories
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.