To clean up a raw transcript, delete the timestamps and the crosstalk, cut the filler words, fix grammar and punctuation, and keep every point each speaker made. When a sentence can’t be saved, cut it. Never let the AI guess. The prompt I paste into ChatGPT or Gemini is below.
Key takeaways
- Clean means readable, not shorter. Every point the speaker made survives.
- Tell the model it may not condense. Left alone, it summarizes.
- Label every speaker the same way every time before you edit a word.
- Cut sentences you can’t recover. Guessing puts words in someone’s mouth.
- Flag quotes and numbers as you clean. That list becomes your article.
What does cleaning up a raw transcript mean?
Cleaning up a raw transcript means turning machine output into text a person could publish as is, with the speaker’s meaning intact. Meeting and podcast recording tools often hand you the same mess: timestamps, “um,” false starts, and one speaker split across five turns.
The trap I watch for is the silent summary. You ask a model to “clean this up,” and it hands back something shorter that reads well. It can also drop the speaker’s best example or an objection a guest raised. Nothing in the output tells you what went missing.
A clean transcript keeps the argument and drops the noise. Grammar is the easy part.
How do you clean up a raw transcript step by step?
Work in this order. Each step makes the next one easier.
- Build a speaker roster. List each speaker, their role, and whether you can quote them by name. If the file says “Speaker 1,” find out who that is. Never invent a name or a title.
- Standardize the labels. Use one bold label per speaker, spelled the same way every time. Merge split turns from the same person into one block.
- Delete timestamps. Park them in a separate list if you need them for a video description.
- Cut the fillers. Remove “um,” “uh,” “like,” “you know,” “basically,” “I mean,” “kind of,” “sort of,” and repeated false starts.
- Strip crosstalk and logistics. Audio checks, scheduling and side conversations go.
- Repair sentences. Fix grammar and finish incomplete thoughts without changing the meaning. If a sentence is unrecoverable, cut it.
- Flag the raw material. Mark every quote, number, customer story and sharp phrase worth reusing. List them at the end.
- Check every number. Keep a statistic attributed to the person who said it, or drop it. A number said on a call is a claim until you verify it.
What prompt cleans up a transcript in ChatGPT or Gemini?
Paste this prompt first, then paste the transcript. For long recordings, paste the transcript in parts and tell the model to wait until the last part arrives.
Objective: Edit this transcript for grammar, spelling and punctuation.
Do not condense it. Keep every point, opinion and example each speaker made.
Steps:
1. Delete all timestamps.
2. Use one bold label per speaker, spelled the same way every time.
Merge consecutive turns from the same speaker.
3. Remove filler words and false starts: "um," "uh," "like," "you know,"
"basically," "I mean," "kind of," "sort of."
4. Remove audio checks, scheduling talk and side conversations.
5. Fix grammar, tense and punctuation. Capitalize proper nouns.
6. Keep each speaker's voice and phrasing. Do not rewrite their opinions.
7. If a sentence is unclear, put it in [brackets] with a note. Do not guess.
8. Keep bold or italics that were in the original.
At the end, list every quote, statistic and customer example worth reusing,
with the speaker who said it.
I will paste the transcript in several parts. Reply "ready" after each part
and start editing only when I say "done."Two changes from my older version of this prompt matter most. The old one asked the model to remove fillers and also keep pauses, and those two instructions fight each other. This one removes fillers and keeps the meaning. The old one also had no list of flagged quotes at the end, and that list is the most useful output of the whole clean.
How do you clean up a Zoom, Teams or interview transcript?
Use the same steps. Fix the speaker labels first, because that’s where meeting tools break most often.
- Zoom and Teams: labels often show display names or “Unknown speaker.” Rename each one to the person’s real name and role before you run the prompt.
- Interviews: the guest is an outside voice. Quote them word for word after cleaning, and never put your own positioning language in their mouth.
- Podcasts: cut the intro music cues and ad reads with the logistics. I use a podcast transcript cleanup and article draft prompt when the episode is headed for the blog.
Should you clean a transcript or condense it?
Clean it first. Condense later, on purpose, in a separate step.
| Clean transcript | Condensed summary | |
|---|---|---|
| Length | Close to the original, minus noise | A fraction of the original |
| Speaker’s words | Kept, with grammar fixed | Paraphrased |
| Quotable | Yes, verbatim | No |
| Best for | Publishing, quotes, article source | Internal notes |
| Risk | Low, if unclear lines get bracketed | Lost points nobody notices |
When you run both in one prompt, you can’t tell which edits were cleanup and which were cuts.
How do you turn a clean transcript into an article?
Use the clean transcript and the flagged list as your source, then write the article as a separate job. I built this process into a skill called transcript-to-article, part of my B2B PMM Skills repo for Claude.
The transcript-to-article skill runs five stages:
- Clean the transcript. The same steps as above, checked against a speaker roster.
- Write the article. It opens with a hook and a direct answer, then key takeaways, the stakes, the mechanism and the proof, with question-based headings.
- Build the distribution package. You get three titles (SEO, provocative, direct), a meta description of 155 characters or fewer, a video description and show notes.
- Create the social assets. It writes a LinkedIn post, a short post under 280 characters, a five-post thread and an image prompt.
- Scan before delivery. It checks every asset for em dashes, banned words, AI tells, unsourced stats, wrong attribution and character limits, with a pass or fail per line.
It reads your brand kit first, so the article uses your positioning, your voice and your calls to action. If the brand kit is missing, the skill stops and asks you to set it up. It won’t invent your positioning. It also won’t promote a number from the recording to a company stat unless you’ve verified it.
Get B2B PMM Skills on GitHub and run transcript-to-article on your next recording.
FAQ
Can ChatGPT clean up a transcript?
Yes. Give it a prompt that forbids condensing, paste the transcript in parts, and check the bracketed lines yourself.
Should I remove filler words from a transcript?
Yes, for anything you publish. Remove “um,” “uh” and false starts. Keep the speaker’s phrasing and opinions as they said them.
Should I keep timestamps?
Delete them from the published text. Keep a separate list if you need timestamps for a video description or show notes.
What do I do with a sentence I can’t understand?
Bracket it and check the audio. If you still can’t recover it, cut it. Don’t let the model guess.
What’s the difference between a clean transcript and a summary?
A clean transcript keeps every point in the speaker’s words. A summary paraphrases and drops material. Make the clean version first.
Scott King helps B2B technology companies build AI marketing systems and create differentiated content.