Interview Synthesizer

What this skill does

Reads 5-10 customer interview transcripts, extracts candidate insights, clusters them by theme, and produces a ranked list. Each insight cites supporting quotes from at least three different interviews and carries a confidence score.

Inputs

• transcripts_dir: directory of interview transcripts (one file per interviewee, named <id>.md).
• Optional interview_guide: the question script used in interviews (helps anchor coding).
• Optional min_support: minimum interview count for an insight to be ranked (default 3).

Steps

1. Read all transcripts. Strip filler (um, you know).
2. Segment each transcript into utterances by interviewee turn. Keep speaker labels.
3. Extract candidate insights using a "code" pass: any utterance that includes a need ("I wish", "I had to"), a workflow ("first I... then..."), or a magnitude ("X times a week"). Save coded utterances with source interviewee_id and timestamp/line.
4. Cluster coded utterances by similarity (TF-IDF + cosine). Drop clusters with unique_interviewees < min_support.
5. For each cluster, write a one-sentence insight statement in the form <persona> <verb> <object> because <driver>.
6. Compute confidence: number of distinct interviewees + recency weight + presence of a magnitude statement. Score 1-5.
7. Pick 2-3 quotes per insight from distinct interviewees. Truncate at one sentence; preserve the speaker tag.
8. Identify outliers: codes mentioned by exactly one interviewee but with strong language; surface in an "Outlier signals" section as candidate hypotheses, not insights.
9. Rank insights by confidence; render a markdown report.
10. Save to synthesis-<topic>.md.

Output

synthesis-<topic>.md with: study header (interviewees, dates), ranked insights (each with statement, confidence, supporting quotes, source files), an "Outlier signals" section, and a "Methodology" footer. Plus a CSV insights.csv for tracking.

Verification

For each top insight, confirm at least 3 distinct interviewees are cited; if a quote is reused from one interview to support multiple insights, mark it but discount confidence. Re-cluster with a different similarity threshold and confirm top insights remain stable. Manually code 10% of utterances and compare against the auto-coding for sanity.

Edge cases

• Transcripts with one dominant talker (researcher monologuing): exclude their utterances from coding; only interviewee speech counts.
• Very short interviews (under 15 minutes): mark as "limited signal" and lower confidence.
• Interviews on sensitive topics (health, finance): preserve consent; never include identifying quotes without anonymization.
• Mixed languages across interviews: cluster within each language separately; cross-language synthesis is unreliable here.