How to Cut a Customer Interview Video to 90 Seconds
Make this article actionable
Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.
The interview ran forty-seven minutes. The clip has to be ninety seconds, sit above a pricing table, and survive a buyer who has already read three competitor case studies this week. Most teams answer that brief by pulling the four most emotional moments out of the transcript and stitching them together. The result looks like a testimonial and says nothing: a customer smiling, a logo, a generic sentence about "great partnership," no claim anyone could check.
A 90-second customer interview video works when you build it backwards from one claim the customer can defend with their own records. That changes the order of work: transcript first, claim inventory second, edit third, approvals last.
Decide the one claim before you touch the timeline
Write the claim as a sentence a buyer could argue with. "Onboarding dropped from about six weeks to nine days" is a claim. "Working with the team has been transformative" is a mood.
Ninety seconds is roughly 225 spoken words at a normal interview pace of about 150 words per minute. That budget is too tight to carry two claims. Pick one and let the rest of the interview go. If the viewer can repeat the claim after one watch, the cut is working; if they can only repeat the feeling, you have made a brand film, not a customer story.
Turn the useful parts into next steps
Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.
Step 1: Record for the edit, and get the release signed first
The release is not paperwork to collect afterwards. Before recording, get written permission covering marketing use of video and stills, territories, duration, whether you may edit and shorten, whether name, face and logo may appear, and how the customer can withdraw. Keep the signed copy with the raw file.
- Ask each remote participant to record locally if the tool allows it, and keep both the local audio and the call recording.
- Use one microphone per speaker, 48 kHz, and silence notifications.
- Capture ten seconds of room tone; you will need it to fill gaps after cuts.
- Record a sixty-second warm-up you will never use. People settle into their voice around minute two.
Step 2: Transcribe with timestamps, and treat the text as the master
Get a transcript with word-level timestamps and export it as plain text and subtitle files, named by client, topic, date and version. The transcript, not the editing app, is your master document, because the edit will be rebuilt at least once.
LLM Gateway's release tracker, updated November 20, 2026, reports that six new AI models were released in the last seven days. Tools rotate faster than the videos they help produce: if the only copy of your cut lives in one app's project format, a pricing change or shutdown can strand it.
The same caution applies to the transcript. If you paste an interview into an assistant to find the strongest quotes, check what happens to that conversation. Google said on November 16, 2026 that new Gemini settings let users keep chat history while opting out of generative AI model training, and that with the new History setting off, future chats are saved for around two months. Decide deliberately whether a customer's unscripted words belong in that kind of chat, and redact names, account numbers and pricing before you paste.
Step 3: Build a claim inventory from the transcript
Read the transcript once and mark every place the customer describes the problem, the trigger, the before-state, what changed, a number and a caveat. Then put those in a table.
| Beat | What the customer said | Evidence | Status |
|---|---|---|---|
Problem | "Every renewal was a scavenger hunt." | Their own words | Usable |
Trigger | "We lost a customer over a missed date." | Their support ticket count | Needs confirmation |
Before-state | "Two people, four spreadsheets." | Screenshot of the old process | Usable |
The change | "Now it's one queue." | Product screen recording | Usable |
Number | "We cut the cycle from six weeks to nine days." | Their dashboard export | Needs date range |
Caveat | "It took us a month to migrate." | Their project notes | Keep it |
Two rules: no number enters the script without evidence attached to its row, and a caveat is not a weakness — it is what makes the number believable to someone who has been sold to before.
Step 4: Verify every number with the customer, in writing
Send a one-page verification note: the sentence you intend to use, the number, its source and its date range. Ask for a one-line reply confirming or correcting it, and keep that reply with the release.
This is where AI assistance is most dangerous and most useful. A model will happily smooth "six weeks to nine days" into "a 78% reduction" — arithmetic on a number nobody verified, at a precision the customer never claimed. Microsoft's AI Transparency Initiative, announced November 17, 2026 with more than 40 organizations including Accenture, Adobe, KPMG, Mastercard and Volvo Cars, sets out a principle worth borrowing: organizations should be able to see, verify and control what an AI system did, and "agents cannot be the sole source of truth about their own behavior." That is about AI systems, not interviews; my reading is that the discipline transfers. An assistant's paraphrase of a dashboard is not the record — the customer's export is.
Where evidence is thin, keep the qualitative version: "renewals stopped slipping" is defensible without a metric.
Step 5: Script the 90 seconds as six beats
Write the script as timecodes and word counts before you open the editor.
| Time | Beat | Budget |
|---|---|---|
0:00–0:06 | The problem, in the customer's words | 15 words |
0:06–0:25 | The before-state, with one concrete detail | 45 words |
0:25–0:50 | What changed and who did what | 60 words |
0:50–1:10 | The verified number and how it was measured | 50 words |
1:10–1:25 | What they would tell a peer | 40 words |
1:25–1:30 | Name, role, company card | 15 words or silent |
The number arrives late, after the viewer has a reason to care, and the customer gets the last word before the card.
Step 6: Cut soundbites, not sentences
Choose three or four bites of twelve to eighteen seconds each. A usable bite starts on a verb, ends on a full thought, and has no throat-clearing at the head.
- Trim "yeah, so, I mean" from the front of a bite; leave the rest intact.
- Never reorder words inside a bite. If it needs reordering, it is a paraphrase, and a paraphrase delivered in the customer's voice is a fabrication.
- Keep one imperfection: a single hesitation reads as real.
- Log every in and out point in a cut list, so the edit can be reproduced or handed over.
Step 7: Build the spine, then cover the cuts
Lay the interview audio as a continuous spine and cut picture separately, letting audio lead into each shot and trail out of the previous one, so the viewer hears a conversation rather than clips.
- Leave one to two seconds of handle on every clip for caption and music changes.
- Avoid two cuts closer than two seconds apart in the same framing. Use a slight punch-in or a cutaway instead.
- Cover transitions with b-roll, never a claim. An unrelated stock shot under "we cut the cycle to nine days" invites the viewer to distrust the number.
- Blur names, emails, ticket IDs and internal URLs in every screen recording, and check that any third-party logo in frame is one you may show.
Step 8: Captions, loudness, and the vertical version
Most viewers meet this video muted, in a feed, at arm's length.
- Ship burned-in captions and a sidecar subtitle file: two lines maximum, around 32 characters per line, one to two seconds per card.
- Caption what was said, not a cleaned-up version. If caption and audio disagree, the viewer trusts neither.
- Mix dialogue to about −14 LUFS integrated for web with true peak near −1 dBTP, and keep music under speech.
- Cut the 9:16 version by re-framing, not by letterboxing a 16:9 master.
Step 9: Run one review round, with a deadline
Send a timestamped link, the verification note, and three questions: is anything factually wrong, is anything off-limits publicly, and do you approve your name, face and logo here.
Agree in advance, in the release, that they may veto a sentence but not rewrite the story; otherwise review becomes an editing committee.
Step 10: Export a small, boring matrix
- A 16:9 master at full resolution, plus a compressed review copy.
- 9:16 and 1:1 versions cut from the master, not re-edited.
- File names that carry client, claim, date and version:
acme-onboarding-9days-2026-11-22-v3.mp4. - One archive folder per customer: raw, transcript, cut list, release, verification note, exports.
Checklist before it goes on the page
- One claim, stated in a sentence a buyer could argue with.
- Every number traced to a dated export and confirmed in writing.
- Release signed, covering edit rights, logo, territory and duration.
- Caveat included where the customer volunteered one.
- No word reordered inside a soundbite.
- No claim covered by unrelated b-roll.
- Names, emails and account data blurred in every screen recording.
- Captions match the audio word for word.
- Customer approved the final cut in writing.
- Master, transcript, cut list and release archived together.
Common mistakes
Building a highlight reel and calling it a story. Four emotional moments without a claim leave a feeling and no argument.
Letting the assistant do the arithmetic. Rounding and "3x faster" claims are where invented precision enters; do the math from the customer's numbers or drop it.
Putting music under the number. The one sentence carrying the argument should be the cleanest audio in the cut.
Why the transcript-first order holds up
Model names change monthly and features you built around can be retired. What survives is what you own as files: the raw recording, the timestamped transcript, the cut list, the verification note and the release.
That is also the cheapest way to make a second and third video from the same interview: the claim inventory shows which other beats are already verified, and the cut list shows where they sit. One good forty-seven-minute conversation is usually three videos — as long as you wrote down what was true before you started cutting.

