---
name: interview-batch-synthesis
description: Synthesize a batch of interview transcripts into JTBD statements and themes, every quote verbatim and sourced. Use this skill when: New batch or biweekly — Fires once transcripts cross the threshold, or on schedule.
---

# Synthesize interview transcripts into JTBD themes

Goal: synthesize a batch of interview transcripts into JTBD statements and contradictions.

Connected tools this skill uses: drive, notion, slack.

## When to use

- New batch or biweekly — Fires once transcripts cross the threshold, or on schedule.
- Or when the user asks directly to synthesize a batch of interview transcripts into JTBD statements and contradictions.

## Method

Run the interview batch synthesis.

1. List every new transcript in {{interviews_folder}} since the last synthesis
   run ({{last_run_date}}), and confirm the batch has reached
   {{min_batch_size}}; if not, stop and wait for the next sweep.
2. For each transcript, extract job-to-be-done statements and pains, with one
   verbatim quote per finding — speaker, timestamp, and date, never
   paraphrased.
3. Roll the batch up into themes across interviews, noting how many distinct
   interviews support each theme.
4. Cross-check each theme against the synthesis memory from past batches;
   flag anything that contradicts a prior finding instead of silently
   overwriting it.
5. Validation: confirm every quote in the output matches the source
   transcript's exact wording, speaker, and timestamp before it ships.
6. Patch the research-insights view with the new themes and contradictions,
   save the batch report to `reports/research/{{date}}-synthesis.md`, and
   post the synthesis summary to {{delivery_channel}}.

Never share interview content or quotes outside the team's surfaces — this
synthesis stays internal.

## JTBD extraction

Pull job-to-be-done statements and pains from raw transcript text, one per real signal.

## Method
1. Read the full transcript, not just the summary or highlights.
2. Identify moments where the interviewee describes a goal, a workaround, or a frustration in their own words.
3. Phrase each as a JTBD statement: "when [situation], I want to [motivation], so I can [outcome]."
4. Separately, capture each distinct pain as its own line, not folded into the JTBD statement.
5. Tag each JTBD and pain with the persona it came from, if known.
6. Attach the supporting verbatim quote to each extracted statement (see the verbatim-quote policy).
7. Discard a candidate JTBD that has no supporting line in the transcript — do not infer one from context alone.

## Rules
- One JTBD statement per distinct goal; do not merge two different goals into one statement for brevity.
- Every JTBD and pain must trace to a specific line in the transcript.
- Do not extract a JTBD from a hypothetical or a leading question the interviewer asked — only from what the interviewee actually said.
- Persona tags follow the team's stored persona list; do not invent a new persona name mid-run.

## Output
Per transcript: a list of JTBD statements and a list of pains, each with persona tag and supporting quote reference.

## Verbatim-quote policy

Attach one exact, sourced quote per finding — speaker, timestamp, and date — never a paraphrase.

## Method
1. For each JTBD or pain finding, locate the transcript line that best supports it.
2. Copy the line exactly as transcribed — no smoothing, no summarizing, no combining lines.
3. Attach the speaker's name, the timestamp in the transcript, and the interview date.
4. Re-read the quote against the transcript once more before it leaves the draft; any mismatch fails the check.
5. If no line reads cleanly as a standalone quote, drop the quote and keep the finding without one.
6. Carry the same quote forward unchanged if it's cited again in a later theme rollup.

## Rules
- A quote is verbatim, with speaker, timestamp, and date, or it is not a quote — never paraphrase-as-quote.
- Every quote traces back to one specific transcript and one specific moment in it.
- A missing quote never blocks a finding from shipping — the finding stands on its own without one.
- Quotes stay internal; this policy governs sourcing, not distribution (see the recipe's own guardrail on sharing).

## Output
Per finding: either one verbatim quote (speaker, timestamp, date) or no quote field, with the finding still recorded.

## Contradiction flagging

Compare this batch's findings against stored synthesis memory and surface disagreements instead of overwriting them.

## Method
1. Load the running JTBD/theme inventory from memory before rolling up this batch.
2. For each new theme or JTBD, check whether a prior synthesis made a claim that conflicts with it.
3. If it reinforces a prior finding, increment its supporting-interview count in memory.
4. If it conflicts, do not silently replace the prior finding — record both, with their respective evidence and dates.
5. Write a short contradiction note explaining what changed: new segment, new context, or a genuine reversal.
6. Leave the resolution of a contradiction to the PM — this skill surfaces it, it does not decide which finding wins.

## Rules
- Never overwrite a prior synthesis finding without a flagged contradiction note — silent overwrites are a hard fail.
- A contradiction needs its own supporting quote, same verbatim standard as any other finding.
- Low-confidence disagreement (one interview against many) is still flagged, just labeled with its weight of evidence.
- Contradictions accumulate in memory across batches; do not delete a past contradiction once logged, even after it's resolved.

## Output
A contradictions list: each entry names the prior finding, the new finding, supporting evidence for both, and a status (open/resolved).

## Recurring jobs

This skill pairs with scheduled jobs. Offer to set each one up; only create the ones the user approves:

- Biweekly interview batch synthesis (Every other Monday at 6:00 AM, or as soon as the transcript batch threshold is reached) — https://eluu.ai/cookbook/interview-batch-synthesis/jobs-setup/biweekly-interview-batch-synthesis.md