A Discovery Call Synthesis Prompt That Grades the Evidence, Not the Vibes

Sales and product teams run dozens of discovery calls a quarter, then synthesize them from memory the night before the roadmap meeting. The loud interview wins. A vivid quote from one account becomes "customers are telling us," and a build decision rides on a pattern that was really one person. A discovery call synthesis prompt fixes the honest part: it reads the transcripts, pulls claims with the exact quote behind each, counts themes by distinct participants, and grades how solid the evidence actually is.

This is not the same job as recapping a call. Recap compresses one conversation. Synthesis is the harder, higher-stakes work of turning a batch of calls into something you'd defend in a readout.

Where the ranking pages stop

The pages that rank for discovery-call prompts are built for a different step. Allego's guide offers a clean post-call summarization prompt but, by its own shape, "does not extract verbatim quotes, cluster themes across multiple calls, or grade evidence quality" — it's a single-call summary inside a broader ChatGPT-for-sales overview (allego.com). Arrows publishes a genuinely useful template that analyzes your ten most recent transcripts for missed qualification, but it "does not include locked output contracts, verbatim quote extraction, cross-call theme clustering, or evidence strength grading," and it funnels to a product demo (arrows.to).

So the gap is specific and consistent. No verbatim evidence trail. No frequency count by distinct participant. No grade separating a solid finding from a single vivid anecdote. That's exactly where a synthesis prompt earns its keep.

What synthesis actually produces

The prompt turns transcripts into a decision-ready structure, not a wall of bullets. Across a batch you get:

• Atomic claims — one idea per row, each with the verbatim quote it came from
• A behavior-vs-opinion label — what someone did versus what they said they'd do
• Themes counted by distinct participants, so one talkative account can't fake a pattern
• A severity grade anchored to real cost, not gut feel
• Contradiction flags where two participants or segments disagree
• Evidence grades — solid, promising, or thin — on every theme

The behavior-vs-opinion split does the heaviest lifting. "I'd definitely pay for that" is an opinion; "we built a spreadsheet to work around it" is behavior. Weighting them the same is how discovery lies to you. A good prompt keeps them in separate columns.

Anatomy of the prompt

Two inputs per call, and the output contract goes last so it survives a long transcript paste.

Variables
  {{transcript}}        → one interview, verbatim or noted
  {{participant}}       → role, segment, company size
  {{research_focus}}    → the question this round is trying to answer

Prompt
  Role: a discovery analyst who separates evidence from opinion.
  Task: extract atomic claims from {{transcript}}; attach the
        verbatim quote and a behavior/opinion/hypothetical label
        to each; flag leading questions and coverage gaps.
  Constraints: never paraphrase a quote into a claim without marking it;
               if a claim rests on a hypothetical, label it low-confidence;
               do not merge two participants' words into one claim.

Output contract (last lines of the prompt)
  ## Claims
  | Claim | Verbatim quote | Type | Confidence |

  ## Coverage gaps
  - <topic the research_focus needed but the call skipped>

  ## Contradictions
  - <where this participant undercuts themselves>

Run the core prompt per call, then feed the stacked claim tables into a clustering pass with the participant roster. That second pass is where frequency-by-distinct-participant and severity grading happen — and where one loud voice stops becoming a fake trend.

Step-by-step usage

1. Extract claims one call at a time

Paste a single {{transcript}} with the {{participant}} context. Running per call keeps quotes attached to the right person and stops the model from blending two accounts into one imagined customer.

2. Label behavior against opinion

Check the Type column. If everything came back "opinion," the prompt didn't find evidence of action — which is itself a finding about how you're running discovery. Real behavior is the stuff worth building on.

3. Cluster across the batch

Feed all the claim tables plus the roster into the clustering pass. Themes come back ranked by distinct participants, not raw mentions. Claude holds the distinct-participant count reliably when you name it explicitly; GPT-4o will sometimes revert to counting total quotes, so restate "count each theme by unique participant, not by mention" on the last line.

4. Read the contradictions

Contradictions are findings, not noise. If enterprise buyers want governance and self-serve users want speed, that tension belongs in the readout as a decision, not averaged into a mushy middle.

5. Grade before you present

Every theme carries a solid / promising / thin grade. A thin theme isn't wrong — it's under-evidenced, and it should ship with the one question that would firm it up next round rather than a confident recommendation.

Prompt-craft patterns worth stealing

Quote-first, claim-second. Force the verbatim quote into the same row as the claim and paraphrase drift can't hide. If there's no quote, there's no claim — the model has to admit it inferred. That traceability is what lets a skeptical exec follow any finding back to the exact line someone said.

Every claim row must carry the exact words from {{transcript}}.
If you cannot quote it, do not assert it — write [inferred] instead.

Confidence as a column, not a caveat. Burying uncertainty in prose lets it evaporate by the time it reaches a slide. A Confidence column that travels with each claim keeps the hedge attached to the data all the way to the readout.

Watch-only for single-call signals. A vivid one-off shouldn't be deleted or promoted. Mark it watch-only so it's on the radar for next round without pretending it's a pattern yet. Most teams do the opposite: they either forget it or over-index on it because it was memorable.

One caveat worth stating plainly: synthesis can't rescue bad discovery. If your calls are pitches with a few questions bolted on, the prompt will faithfully report that you gathered opinions, not evidence. That's useful feedback, but it's about the calls, not the tool.

Variables you'll set

Getting started

1. Gather your transcripts, verbatim where you have them.
2. Run the claim extractor on each call with its {{participant}} context.
3. Check the behavior-vs-opinion split before trusting any theme.
4. Cluster all the claim tables with the roster to get distinct-participant counts.
5. Read the contradictions as decisions, not noise.
6. Grade each theme solid / promising / thin before it reaches a slide.
7. Present with the quote trail intact, so every claim survives a follow-up question.

Once the calls are synthesized, a MEDDPICC qualification prompt scores whether the deal behind them is real before you forecast it, and a sales follow-up email prompt handles the next touch. Browse the full set of prompt packs when discovery is only one part of the workflow you're wiring.