Expert Interview Synthesis

Fifteen transcripts, read weeks apart, become a vague impression plus the two quotes you happened to highlight. This reads them side by side and produces the thing you actually paid for: what the interviewees agree on, what they genuinely disagree about and whose vantage point makes them credible on that specific question, every number anyone cited, how the views moved over time, and what nobody has answered yet.

Why this skill exists

Interviews are expensive — in fees, in the interviewer's time, or in the goodwill of the customers and references who gave an hour — and their value decays in a specific way. Each conversation is useful when you have it, but the insight lives in the pile rather than in any single transcript, and the pile is never read as a pile. So the researcher, product manager, or deal lead carries an impression into the debate, and the impression is weighted toward whoever they spoke with most recently.

The disagreements are the worst casualty. When a former head of sales and a current channel partner contradict each other on churn, or when two references describe the same candidate's judgment in opposite terms, that contradiction usually is the question. But it only becomes visible when the transcripts are side by side, and it usually reads as noise when the calls are three weeks apart. The same goes for view drift: users who were enthusiastic about a feature in January and hedged by May are telling you something, and nobody notices because nobody re-reads January.

The last output matters most for spend. Knowing which questions remain unanswered means the next interview goes at the gap instead of re-covering ground you already bought.

Inputs

Required:

• Two or more interview transcripts on the same company, product, market, candidate, or question. Five to twenty is the intended range; below five the synthesis is thin, above twenty it still works but flag that coverage may be uneven.

Optional but high-value:

• Your current hypothesis, in a sentence or two — this enables the contradiction check, which is the section most likely to change your mind
• Each interviewee's vantage point if not clear from the transcript: role, organization, seniority, function, and how recently they held the role or used the product
• Interview dates, if not in the transcripts — required for the view-drift analysis
• The specific questions you are trying to answer
• Your plan's or model's key assumptions, so quantitative data points can be tested against them
• Prior synthesis on the same subject, for a what-changed comparison
• The context — product discovery, vendor evaluation, hiring references, market study, live transaction — because it changes which questions matter and which confidentiality limits apply

Missing-input behavior: Transcripts alone are enough to run. Without a stated hypothesis, skip the contradiction section and say so — do not infer a hypothesis from the transcripts and then test it against them, which is circular. Without interview dates, skip the view-drift section explicitly rather than guessing at chronology from content. If the transcripts are on materially different subjects, say so and either scope to the largest coherent group or ask which subject to synthesize.

How it works

1. Build the interviewee roster first. For each transcript: date, the person's role and organization, function, seniority, how recently they held the role or used the product, and their relationship to the subject (former employee, current or former customer or user, competitor, supplier, channel partner, industry consultant, reference, internal stakeholder). Every claim downstream is attributed to a row in this table, because a claim without a vantage point cannot be weighted.
2. Extract claims, not summaries. Go through each transcript pulling discrete factual and judgmental claims with the supporting quote. Resist compressing at this stage — compression is where disagreements disappear.
3. Cluster claims by topic — demand, pricing, competition, churn and retention, product quality, workflow pain, management or candidate quality, unit economics, regulatory, whatever the interviews actually cover. Let the topics come from the transcripts, not from a template.
4. Identify consensus within each topic, and distinguish real consensus from apparent consensus. Three former employees of the same company from the same era are one data point with three voices, not three independent confirmations. Say so when it happens.
5. Identify genuine disagreements. For each: state both positions, who holds each, quote both, and assess which vantage point is more credible on that specific question — a former CFO is authoritative on margin structure and no better than anyone else on competitive win rates; a candidate's former manager is authoritative on how they take feedback and no better than anyone else on how they lead peers.
6. Table every quantitative data point — market share, pricing, growth, churn, unit costs, cycle times, capacity, adoption rates, time-to-value — with the interviewee, their vantage point, the date, and the quote. Note where two people give different numbers for the same metric.
7. Trace view drift across the date range where dates are available: which views strengthened, which softened, and whether the change tracks something observable in the business or product.
8. Test against the stated hypothesis. Flag every place the evidence directly contradicts it, and be direct — this is the section with the most value and the most temptation to soften.
9. Assess coverage and bias in the sample. Which vantage points are over-represented, which are absent, whether the interviews were sourced in a way that could have selected for a view (references chosen by the candidate, customers introduced by the vendor, experts recruited on a keyword), and how much weight the whole body of evidence deserves.
10. Write the next-interview list from the unanswered questions, specifying the vantage point that would answer each.

Output format

# Interview Synthesis — [company / product / market / candidate]
Interviews: [N] | Date range: [earliest]–[latest] | Sources: [expert network / discovery / references / internal]
Hypothesis tested: [stated / none provided]

## What the interviews establish (one paragraph)
The two or three things this body of evidence actually supports, and the
confidence they deserve.

## Interviewee roster
| # | Date | Role | Organization | Vantage point | Recency | Authoritative on |
|---|---|---|---|---|---|---|

## Consensus
### [Topic]
**View:** [ ]
**Held by:** [interviewee #s] | **Independent sources:** [count — not voice count]
> "[strongest supporting quote]" — Interviewee [#], [vantage point]
**Confidence:** [high / medium / low] — [why, referencing independence and vantage]

## Disagreements
### [Topic] — the question at issue: [ ]
**Position A:** [ ] — Interviewee [#s]
> "[quote]" — Interviewee [#], [vantage point]
**Position B:** [ ] — Interviewee [#s]
> "[quote]" — Interviewee [#], [vantage point]
**Whose vantage point is stronger here:** [assessment specific to this question]
**What would settle it:** [ ]

## Quantitative data points
| Metric | Value | Interviewee | Vantage point | Date | Quote |
|---|---|---|---|---|---|
Conflicting figures for the same metric: [flagged explicitly]

## How views shifted over the date range
[Only if dates available. What strengthened, what softened, and whether it
tracks anything observable.]

## Against the hypothesis
[Only if a hypothesis was provided. Every direct contradiction, quoted, with
the interviewee's vantage point. Be direct.]

## Coverage assessment
Vantage points represented: [ ] | Missing: [ ]
Over-representation or sourcing bias: [ ]
Overall weight this evidence deserves: [ ]

## Next interview list
| # | Open question | Vantage point that would answer it | Why it matters |
|---|---|---|---|

Weighting by vantage point

Credibility is question-specific, not person-specific. The useful test is always: was this person in a position to know this particular thing?

• Former employees — authoritative on internal process, cost structure, and culture; decay quickly on current performance. Note how long ago they left and at what level. A former SVP two years out knows the org, not this quarter.
• Current and former customers or users — the best source on why they buy, switching costs, workflow pain, and pricing pressure; representative only of their segment. One enterprise customer says nothing about small-business churn; one power user says nothing about onboarding.
• Competitors — informative on market structure and competitive dynamics, systematically biased on the subject's position and win rates.
• Suppliers and channel partners — good on volumes, lead times, and order patterns; often the best quantitative source; usually see only their slice.
• Industry consultants and analysts — broad view, frequently the least specific, and sometimes recycling what they heard on other calls including yours.
• References — chosen by the person being assessed, so the sample is selected; still authoritative on the specific situations they observed firsthand. Weight what they saw over what they conclude.
• Internal stakeholders — closest to the mechanism and furthest from neutrality; note whose budget, headcount, or reputation the answer affects.
• Seniority cuts both ways — executives see strategy and are further from operational reality; individual contributors see the mechanism and not the aggregate.
• Recency beats seniority for anything about current conditions.
• Independence matters more than count — several people from the same organization and era are one observation. Weight by independent vantage points, not by headcount.

Guardrails

• Every claim is attributed to a numbered interviewee with their vantage point. An unattributed claim in a synthesis is worse than useless, because it inherits false authority from the aggregate.
• Quote exactly. Verbatim, with the interviewee number. No paraphrase presented as a quote, and no composite quotes assembled from multiple speakers.
• Never merge disagreeing interviewees into a single view. Averaging away a contradiction destroys the most valuable content in the pile. If two people disagree, the output shows both.
• Distinguish independent confirmation from repetition. Say explicitly when apparent consensus comes from a single vantage point echoed by several voices.
• Do not fill gaps with general industry knowledge. If the interviewees did not address something, it belongs in the next-interview list, not in the synthesis dressed as a finding.
• Interview testimony is opinion. These are individuals' recollections and judgments, frequently about organizations they no longer work for or products they used months ago, sometimes wrong and sometimes stale. The synthesis reports what was said, weighted — it does not convert testimony into fact.
• Confidential and restricted information. Interviews can surface content the speaker was not entitled to share: an employer's undisclosed results or plans, a specific unannounced customer win or loss, a pending transaction, a former colleague's private circumstances, or anything the speaker describes as confidential to their current organization. Flag it. Do not incorporate flagged content into the synthesis; surface it separately for whoever owns compliance, legal, or research ethics on the user's side. Where an interviewee appears to be a current employee of the subject organization discussing confidential specifics, say so prominently. In regulated contexts such as public-company research, this content may carry legal consequences; this skill does not replace the organization's expert-network policy, chaperone requirements, pre-clearance process, or research-consent rules.
• Not a recommendation. The synthesis maps what is known, contested, and unasked. It does not decide the hire, the purchase, the product bet, or the investment.
• Licensing and privacy: expert network transcripts are licensed content restricted to the subscribing organization; customer and user interviews are usually covered by consent terms. Assume an organization-approved AI environment, and do not reproduce transcripts beyond the quotes the synthesis requires.

Quality checks

• Every interviewee on the roster with a vantage point and recency
• Every claim attributed to a numbered interviewee
• Every quote verbatim, no composites
• Consensus sections distinguish independent sources from voice count
• Disagreements preserved with both sides quoted, never averaged
• Credibility assessed per question, not per person
• Every quantitative data point sourced and dated; conflicts flagged
• View drift section run or explicitly skipped for missing dates
• Hypothesis contradictions stated directly, or the section marked skipped
• Coverage gaps and sourcing bias assessed honestly
• Possible confidential or restricted content surfaced separately and excluded from the synthesis body
• Next-interview list specifies the vantage point needed, not just the topic

Example prompts

"Twelve customer discovery interviews on our scheduling product attached, run over the last five months. My hypothesis: clinics will pay for automated reminders because no-shows are their biggest cost. Synthesize them, and be blunt about anywhere the customers contradict me."

"Nine expert calls from a live diligence process on a specialty distribution business. Synthesize by topic, table every number anyone gave us, and tell me what we still need to ask before the decision meeting in two weeks."

"Six reference calls on our VP Engineering finalist — three she supplied, three we back-channeled. Roll them up, separate what each reference actually observed from what they concluded, and flag where the two groups disagree."

"These eight stakeholder interviews span eighteen months on the same ERP migration. I mainly want the drift — what did people believe early that they don't believe now, and does anything explain it?"

Works well with

Run red-team-premortem on your hypothesis after the synthesis, using the disagreements as ammunition. Pair with presentation-transcript-extractor to test whether leadership's public commentary matches what the interviews say, and market-map-builder when the interviews are sector work rather than single-company work.