Turn customer interview transcripts into case study prompts

To turn an interview transcript into a case study, give the model a fixed Challenge/Solution/Results structure, require line references for every quote and metric, and mark missing facts [NOT IN TRANSCRIPT]. The prompt below produces a grounded first draft, but customer approval and a human check against the recording remain mandatory.

What prompt turns a transcript into a case study?

Copy this block whole, fill the bracketed fields, and paste your transcript between the markers.

You are a B2B content writer drafting a customer case study for [COMPANY], which sells [ONE-LINE PRODUCT DESCRIPTION] to [ICP]. Write in the voice defined here: [PASTE OR REFERENCE VOICE GUIDELINES].

Your only source is the interview transcript between ### TRANSCRIPT START and ### TRANSCRIPT END. Do not use outside knowledge about the customer, the product, or the industry.

### TRANSCRIPT START
[PASTE SPEAKER-LABELED, LINE-NUMBERED TRANSCRIPT HERE]
### TRANSCRIPT END

TASK
Draft a case study of 600 to 900 words with these sections, in this order:
1. Headline: lead with the single most significant quantified result stated in the transcript.
2. Customer snapshot: company and industry, plus the role of the person interviewed, as stated in the transcript.
3. Challenge: what was broken before, why it mattered to the business, and what they had already tried. Include the customer's own words about the pain.
4. Solution: what they implemented, how rollout went, and which capabilities they name.
5. Results: every quantified outcome in the transcript, each with its before and after value where given.
6. Pull quote: one verbatim quote of 15 to 40 words that a prospect would find persuasive.

QUOTE RULES
- Every quotation mark encloses text copied word for word from the transcript. You may cut filler ("um", "you know") and mark omissions with [...]. You may not reorder, paraphrase, or combine sentences from different lines inside one quote.
- After each quote, add the line reference in brackets, e.g. [L42].

METRIC RULES
- Every number, percentage, dollar figure, and time period must appear in the transcript. Copy it exactly and add its line reference.
- Do not compute new figures from transcript numbers unless the speaker stated the calculation.
- If a section needs a fact the transcript does not contain (team size, timeline, before-state baseline), write [NOT IN TRANSCRIPT: describe what is missing] in place of the fact. Never fill the gap with a plausible value.

TONE AND FORMAT
- Plain, specific prose. No superlatives the customer did not use.
- Paragraphs of 2 to 4 sentences. Use the customer's vocabulary for their problem, not our product marketing terms.
- Return the draft, then a separate list titled OPEN ITEMS containing every [NOT IN TRANSCRIPT] flag and every quote you shortened.

The prompt runs on one transcript per session. If you batch several interviews, the model can start blending customers, and the line references stop meaning anything.

What is each part of the prompt doing?

Five elements carry the prompt:

• Role: Who is writing and for whom.
• Context: The source material and durable artifacts you attach.
• Task: The work the model must complete.
• Section schema: The required order and contents.
• Constraints: The rules for quotes and output.

Knowing which line does what lets you change one without breaking the others.

How do role and context constrain the draft?

The role line and the voice reference tell the model who is writing and for whom. Most adaptation happens when you swap in your own company description and ICP, then attach your voice guidelines.

It should not happen inside the prompt each time. Your voice guidelines and company profile are durable context artifacts you build once and attach to every session. Your audience persona belongs in that durable context too. The prompt is the invocation point for that context, not the place it lives.

The “only source” clause sits in this block on purpose. Restricting the model to provided documents, alongside giving the model explicit permission to say it does not know, reduces hallucinations. A model that knows your product category from training data may otherwise pad the Solution section with features the customer never mentioned.

How do task and structure shape the output?

The six-section schema is a convention, not a standard. Challenge, Solution, and Results show up under slightly different labels across trade publications, and the Challenge section should carry the customer’s fears and frustrations rather than a neutral problem statement. That is why the prompt asks for “why it mattered to the business” and the customer’s own words about the pain.

The headline instruction inverts the usual order by leading with the result. Prospects skim, and a quantified outcome in the first line does more work than a company name. The Results rule forcing a before-and-after pair exists because a lone after-state number (“now under 5%”) is unpersuasive without its baseline. Change the section count if your template differs, but keep the pull quote as its own section. Otherwise the model tends to bury the best line in paragraph four.

How do constraints keep the output verifiable?

The quote and metric rules are the fabrication guardrails, and the line-reference requirement makes them checkable. A reviewer can open the transcript at [L42] and confirm the words in seconds. Without references, verification means rereading the whole call.

The [NOT IN TRANSCRIPT] instruction gives the model an approved way to fail. The marketer can use the OPEN ITEMS list at the end to draft a follow-up email to the customer. In practice, the marketer often gets the most value from that list on the first pass because it identifies the numbers still needed before the draft can go to approval.

How should you prepare the transcript first?

A clean, speaker-labeled, line-numbered transcript gives the model stronger source material than an unedited export. Do this before you paste anything.

Rename speakers from “Speaker 1” and “Speaker 2” to name and role, because the prompt asks the model to identify the interviewee’s title and to attribute quotes. Add line numbers if your transcription tool does not export them. Then cut the scheduling chatter at the top and the goodbye at the bottom. Leave the disfluencies in the body, since the prompt handles them and a raw transcript is easier to verify against the recording than an edited one.

What should you redact and approve?

Interview recordings and their transcripts are personal data under Article 4 of GDPR, and California classifies audio and transcripts as personal information under CCPA. Where you paste them matters.

On consumer ChatGPT tiers, OpenAI may use content for training by default unless you switch off the “Improve the model for everyone” control, and clicking thumbs up or down can pull the whole conversation into training even after you opt out. Business and Enterprise plans do not train on your data by default and come with a Data Processing Agreement. Standardize the team on a business or API tier for anything client-identifiable.

Before pasting, strip these from the transcript:

• Third parties named in passing: Colleagues, competitors, and vendors who did not consent to be in a case study.
• Contract terms: Remove dollar values and seat counts that sit under NDA. Remove protected renewal dates too. Keep the outcome metrics the customer stated, and decide later whether to publish them as ranges.
• Unreleased product detail: Roadmap items the customer mentioned from a briefing.
• Personal remarks: Anything about individuals’ performance, health, or employment status.

For sign-off, send one email after the draft exists: “Attached is the draft built from our call on [date]. Reply ‘approved’ or return edits in track changes by [date]. Nothing publishes without your written OK.” Ask in the interview itself who has authority to approve, since the person who signed the MSA rarely remembers the publicity clause.

Where approval stalls, offer a named sales-only version or an anonymous version that keeps the metrics. Anonymizing the customer while keeping the specific figures preserves most of the asset’s value and can reduce logo and legal review.

How do you go from transcript to first draft in four prompts?

The single prompt above handles the job, but splitting it into four passes separates extraction from prose generation. For long documents, quote-first grounding extracts the relevant passages word for word before the model does anything with them.

Start with an extraction pass. Paste the transcript and ask for a fact list only. Ask for every metric and quotable sentence with line references.

Include every named capability and a separate list of before-state statements. No prose. Save this list. It becomes the source of truth for every downstream asset, and the drafting passes read from it rather than from the raw transcript.

From that list, ask the model to propose two or three narrative arcs and name the headline metric for each. A transcript often contains one operational win, such as tickets or hours, and one commercial win, such as renewals or revenue. Choosing the lead is a judgment call the model should surface rather than make. This is the part most teams skip, and it is why their case studies all open on the same generic productivity claim.

Then run the full prompt with your chosen arc pasted under TASK as a one-line steer: “Lead with the retention outcome.” Keep all the quote and metric rules.

Finally, run a tightening pass on the draft alone, with the transcript out of the context window, asking only for sentence-level edits. Request shorter paragraphs and banned-term removal, then ask the model to swap product jargon for the customer’s phrasing. The tightening pass is forbidden from touching anything inside quotation marks or brackets.

The Vault has the template that operationalizes this. The Prompting Playbook is ready to plug into your existing stack.

What does a grounded output look like?

Judge the output on a constructed sample. The excerpt below is illustrative and uses invented names, but the mess is realistic: two customer speakers deal with a number that changes mid-sentence before one customer declines to answer a question.

[L01] Interviewer: Before you brought us in, what did the support queue look like?
[L02] Dana (VP Customer Ops): A mess. We had, um, three people triaging by hand.
[L03] Dana: About 40 percent of tickets sat past 24 hours. Our SLA was 24.
[L04] Priya (Support Lead): It was 38, 40, somewhere in there. Q3 last year.
[L05] Interviewer: What did that cost you?
[L06] Dana: Renewals. We lost two enterprise accounts that cited response time in the exit call.
[L07] Priya: And the team was burning out. I had two resignations in one month.
[L08] Interviewer: And after rollout?
[L09] Priya: Tickets over 24 hours are under 5 percent now. That's been stable since February.
[L10] Dana: We also stopped hiring for triage. Those three people moved to onboarding.
[L11] Interviewer: Any number on revenue impact?
[L12] Dana: Not one I'd put my name on yet.

Running the prompt on this excerpt produced the following draft. It is condensed to the sections that show the guardrails working.

Headline: Overdue support tickets fall from about 40% to under 5%

Challenge: Before rollout, three people triaged support tickets by hand [L02], and roughly 38–40% of tickets sat past the 24-hour SLA [L03–L04]. The cost showed up in retention. “We lost two enterprise accounts that cited response time in the exit call.” [L06] The support team was also losing people: “I had two resignations in one month.” [L07]

Solution: [NOT IN TRANSCRIPT: what the customer implemented, the rollout timeline, and the capabilities the customer used. The interviewer moved from before-state to results without covering implementation.]

Results: Tickets over 24 hours dropped to under 5% [L09], and the improvement has held “since February” [L09, year not stated]. The three triage staff moved to onboarding, and the company stopped hiring for triage [L10]. Revenue impact: [NOT IN TRANSCRIPT: customer declined to state a figure, L12].

Pull quote: “We lost two enterprise accounts that cited response time in the exit call.” [L06]

OPEN ITEMS: Solution section missing entirely. Revenue figure not stated. Year for “February” not stated. Before-state percentage given as a range across two speakers.

The model reported the 38-versus-40 discrepancy as a range with both line references instead of picking one. It refused to compute an “87% reduction” figure nobody said. It also flagged the empty Solution section instead of filling it from general knowledge.

How do you stop invented quotes and made-up metrics?

Fabrication in case study drafts comes from two places. The model may paraphrase a quote into something cleaner than the customer said, or it may fill a numeric gap with a figure that sounds right. You address both by forcing a source pointer on every claim and giving the model an explicit alternative to guessing.

The verbatim-or-flag clause does the first job. Reserve quotation marks for exact transcript text and allow shortening only with a visible […] marker. If the model cannot locate a quote, require it to flag the quote rather than smooth it. A related cite and retract pattern requires the model to cite a supporting passage for each claim and withdraw any claim it cannot support.

For metrics, run the sub-prompt below on the finished draft. Attach the transcript before you run it.

For every number, percentage, dollar amount, and time period in the draft below, produce a table with three columns: the figure as written in the draft, the exact transcript line that states it, and a verdict of MATCH, MISMATCH, or NOT FOUND. Do the same for every quotation. Do not correct the draft. Report only.

Anything marked MISMATCH or NOT FOUND is a fabrication or a transcription error, and you resolve it by hand. A human fact-checker should keep the recording open for at least the headline metric and the pull quote. A second model pass catches a lot, and the research on structured verification is encouraging: the Chain-of-Verification method drafts, generates verification questions, answers them independently, and revises. It more than doubled precision on a list-answering benchmark. But a verification pass built on the same transcript inherits the transcript’s own errors, so the recording remains the final authority.

What if the transcript is too long for the context window?

Most single interviews fit without chunking. A 60-minute B2B interview lands around 8,000 to 12,000 words. OpenAI’s rule of thumb is roughly 75 words per 100 tokens, so you are pasting about 10K to 16K tokens plus the prompt.

You need chunking for multi-hour interviews or when you record several stakeholders separately. You may also need it for batch processing across a customer’s whole account history. When you do, never draft from the chunks. Extract from each chunk and draft from the merged fact list.

Split the transcript at speaker turns, never mid-sentence, into pieces of roughly 5,000 words with the original line numbers preserved. Run the extraction pass from the four-prompt sequence on each chunk with the instruction “Report only facts stated in this chunk. Do not summarize or interpret.” Merge the fact lists into one document, dedupe the repeated metrics, and note any figure that changes between chunks.

Then run the drafting prompt with the merged fact list in place of the transcript, and change the “only source” clause to point at the fact list. Line references survive this process because you preserved them at the split. The verification sub-prompt then still works against the original transcript.

Which interview questions make a case study transcript useful?

The prompt cannot extract a before-state baseline that the interviewer never asked for. Most weak case study drafts trace back to a 30-minute call that covered how much the customer likes the product and nothing about the numbers. Ask a stronger set of interview questions so the transcript contains the evidence the draft needs.

Build the call around these questions:

• The baseline: “What was the number before? Tickets, hours, dollars, error rate, whatever you tracked.” Ask for the unit and the time period.
• The cost of the problem: “What did that cost you, in revenue, headcount, or churn?” This produces the Challenge section’s stakes.
• The failed alternative: “What did you try first, and why did it not hold?” This separates your solution from the status quo.
• The rollout: “Walk me through the first 30 days.” Without this, the Solution section comes back empty, as in the worked example.
• The after-state: “Same metric, today. And since when has it held?”
• The peer line: “What would you tell someone in your role at another company who is evaluating this?” This often yields the pull quote.
• The approver: “Who at your company needs to sign off before we publish, and would an anonymous version be easier?”

Results written in the customer’s own words can persuade more than the same numbers in vendor voice. So when a customer gives you a number, follow with “say more about what that meant for the team.” The second answer is often the quote.

How can you reuse the transcript for other assets?

The fact list from the extraction pass is the reusable artifact, and it outlasts the case study. Store it alongside the transcript. Mark the approval status of each metric and quote so every downstream asset can read from it without re-parsing the raw call.

A testimonial is the pull quote plus the headline metric, followed by the customer’s name and title. The prompt is one line: “From the approved fact list, write a 40-word testimonial using only quoted text and the headline figure, with line references.” A LinkedIn post from the customer’s perspective takes the before-state and peer line, then adds one result in the customer’s vocabulary for them to approve and post under their own name. A sales one-pager is the Results section as bullets, the pull quote, and the Challenge in two sentences, formatted for a rep to attach to a follow-up email.

The rule for every derivative asset is the same as for the case study: nothing enters quotation marks that is not in the fact list, and nothing gets published that the customer has not approved in that specific form. Customer approval of a quote for a case study does not automatically cover a paid ad.