How to Write Podcast Show Notes With AI: A Template for Producers

Feed a timestamped transcript to an AI assistant with a fixed template and ask for five parts: a short summary, chapter timestamps, a guest bio, a list of links mentioned, and a flag list of anything uncertain. Then check every name, link and timestamp against the audio before publishing.

The drafting takes minutes. The checking is what makes the notes worth reading, because the three things listeners notice are wrong names, dead or wrong links and timestamps that land a minute away from the topic. Hosting brands publish plenty of show-notes templates; what they do not give you is a drafting prompt that leaves blanks blank, and a check routine that catches what an assistant gets wrong.

Follow me on Instagram@sagnikteaches

The one-page template, with the order that makes AI behave

Save this as a document called "Show notes template" and treat it as the fixed output. The order matters because each part feeds the next: the transcript gives chapters, chapters give the summary, and the summary and chapters together decide which links are needed.

Connect on LinkedInSagnik Bhattacharya

EPISODE TITLE: [working title, under 70 characters]
SUMMARY (2-3 sentences, no hype): ...
IN THIS EPISODE (chapters):
00:00 Intro
[mm:ss] [Chapter title, under 40 characters]
...
GUEST: [name, as they spell it] - [one-sentence bio from their own site or the guest form]
LINKS MENTIONED:
- [thing mentioned] - [URL] - said at [mm:ss]
SPONSOR / HOUSEKEEPING: [only if applicable]
CHECK FLAGS (internal, delete before publishing):
- [uncertain name / figure / link]

Two design choices carry weight. Every link has the time it was said, so you can check it in seconds. And the check flags are a separate block the assistant must fill in, so doubt gets written down rather than smoothed over.

Subscribe on YouTube@sagnikteaches

Start with a transcript you can trust

AI drafts are only as good as the transcript. Most recording and editing tools, and many podcast hosts, can generate a transcript with time codes. Export it with speaker labels and timestamps. If you are choosing a tool, Descript versus Riverside for producers compares two common options.

Fix the names first

Speech-to-text gets personal and brand names wrong, and every downstream draft inherits the mistake. Before you give the transcript to an assistant, do a find-and-replace pass on the names you know: host, guest, company, product, place names within the show. Build a small glossary once per show (the guest's surname exactly as on their booking form, the company written as two words if that is how they write it, product names with their own capital letters) and paste it with every prompt. In an illustrative episode, the transcript spelled the guest's company three different ways in 40 minutes; one pass of find-and-replace fixed all of them in a minute.

Keep what you paste to what the task needs

A transcript can contain a guest's private remarks, an off-the-record aside or a sponsor's unannounced news. Skim for anything marked off the record, and cut it before it goes into any AI tool. If your production company has client rules about unreleased episodes, use the plan your team has approved for that work. Unreleased episodes are embargoed content for some shows, so check before pasting.

Chapters first: the timestamp prompt

Ask for chapters before the summary, because they force the assistant to read the whole episode in order. A prompt that works:

Below is a transcript with time codes in [hh:mm:ss].
Create chapter markers: between 6 and 10 chapters for a 45-minute episode, each starting where the topic genuinely changes.
Rules:
- The first chapter must be 00:00 Intro.
- Use the exact time code of the sentence where the new topic begins; do not round and do not invent times.
- Each title under 40 characters, specific, no clickbait, no numbering.
- After the list, note any chapter where the topic shift was gradual and the time is approximate.
Glossary (spelling): [paste]
Transcript: [paste]

Illustrative output, shortened: "00:00 Intro. 02:14 Why the guest left a hospitality career. 09:40 Pricing a first service package. 17:05 The mistake that cost three clients. 26:30 Hiring a first part-time helper. 38:12 What they would do differently. Note: the move from pricing to the lost clients at 17:05 was gradual; the time is approximate." The last note is the model being honest, which is exactly what you asked for.

Check every timestamp in under ten minutes

Open the audio, click each chapter time and listen for five seconds. Ask two questions: does the new topic start within about five seconds of this point, and does the title describe what you hear? Typical misses are chapters that start one question too early, a time pasted one sentence out, and two topics merged under one heading. If a chapter is off, move the time to the real start and re-title it. For a 45-minute episode with eight chapters this takes about eight minutes, and you will fix one or two chapters on average. The check is the part you must not skip, because a wrong timestamp is the one error a listener hits within seconds.

The format of that list decides whether apps turn it into tappable chapters. Spotify's creator help says it reads chapters from the episode description when each line is a timestamp then a title, such as "(00:00) Intro", the first chapter starts at 00:00, there are at least three, they run in time order and the titles are plain text with no emojis or HTML. It recommends titles under 40 characters and advises against numbering them, because the numbers are removed. Manual chapters show instead of the automatic chapters Spotify generates for some English-language shows, and an edited description can take a few hours to reach every listening platform. YouTube's rule for video chapters is close: the first timestamp at 00:00, at least three in ascending order, and each chapter at least 10 seconds long. Other apps have their own rules, so check the help pages for the apps your listeners use. The time-then-title list in the template above meets both sets of rules.

The summary and the title

With the chapters verified, paste them back with the transcript and ask for the summary. Constrain it:

Write a 2-3 sentence summary for listeners deciding whether to press play.
State who the episode is for and one concrete thing they will learn. Use only claims made in the transcript.
No superlatives, no "deep dive", no "unpack". Then give three title options under 70 characters.

Illustrative output: "A first-time studio owner explains how she priced her first group-class package, the mistake that cost her three clients, and what she changed in her first year of hiring. For small service business owners setting their first prices." Compare it with the audio: did she really say the mistake cost three clients? That figure is a claim; verify it or cut it.

A before and after of a summary

Before, a typical unprompted draft: "In this fascinating episode, we hear the incredible journey of a visionary entrepreneur and the secrets of success." After a constrained prompt and a human edit: "A studio owner walks through setting prices for her first class package, why three clients left, and how she now decides what to charge." The second one is shorter, true, and tells the listener what is inside.

Guest bio and links: where the errors hide

Guest bio

Never ask the assistant to write a bio from the transcript alone. It will fill in titles, employers and achievements from guesswork. Give it the bio the guest supplied (a form field, their site's about page), and ask it to trim to one sentence. In an illustrative case, a draft bio promoted a guest from "owner" to "CEO and founder of an award-winning firm"; the guest had never said "award-winning". The fix is simple: the guest's own words only, and send the final notes to the guest before publishing, if the show's norms allow.

Links mentioned

Ask the assistant to list every product, book, site or person mentioned with the time said, and to leave the URL as MISSING unless the URL appears in the transcript. You then fill the URLs yourself, from the vendor's or author's own site, and click each one. AI-generated URLs are a classic failure: they look right and point to a page that does not exist, or to a different business with a similar name. Never publish a link you have not clicked.

From this transcript list every product, book, website, event or organisation the speakers mention.
Columns: item | who said it | time | exact words used | URL (write MISSING; do not guess).
If a name might be a mishearing, add "uncertain" in a final column.

The "exact words used" column helps the checker: if the transcript said "that pricing book by the woman who also runs the training firm", you know to confirm the title with the guest. That is the part no model can supply reliably.

Claims and numbers: the review that protects the show

Show notes carry your name and the guest's. If notes say "she grew revenue 300 per cent" and she said 30, you have put a false statement under both names. Build a check around claims.

1. Ask the assistant: "List every number, percentage, price, date and named claim in your draft, with the transcript time it came from."
2. For each, listen to the five seconds around that time and confirm it.
3. Delete or soften any claim you cannot confirm. Where a guest made a factual claim you cannot verify, say "she says" rather than stating it as fact.
4. For health, money or legal content, add a line pointing listeners to a qualified professional, and check that nothing in the notes reads as advice the host did not give.

This takes five minutes and is the check I would never drop.

One episode end to end, with times

Take an illustrative weekly interview show run by a two-person production team, 45-minute episodes. Before AI, notes took about 70 minutes: scrubbing through audio for chapters, typing a summary, chasing links. With the template, the flow is:

That is about 38 minutes of which 15 are checking, a saving of about half an hour per episode, 2 hours a month on a weekly show and more on a daily one. Note that the second person step is a rule: the person who produced the notes should not be the only person who checks them. A fresh pair of eyes catches the misspelt guest name the producer has read ten times.

Adjusting the template for different kinds of show

One template does not fit every format, and a prompt that suits an interview show will fail on a solo one. Four adjustments cover most producers.

Solo business show

There is no guest, so drop the bio and add a "Key takeaways" list of three to five bullets, each tied to a chapter. Ask the assistant to quote the host's own words for each takeaway, then keep only the ones the host would actually sign. Solo shows tempt a model to turn a rambling point into a neat principle the host never stated. Illustrative check: the host spent four minutes on why she raised her prices; the draft says "Always raise prices every January". She said "I did it in January." The fix is to quote, not generalise.

Panel or multi-guest show

Speaker attribution gets harder. Give the assistant the speaker labels and ask for "who said what" on any claim or opinion. In notes, name the speaker for a strong opinion ("Guest 2 argues...") so one guest does not appear to endorse another's view. If the transcript's speaker labels are unreliable, which happens when voices overlap, label the key quotes by hand.

Video podcast

If the episode goes on a video platform as well, the same chapter list doubles as YouTube's video chapters, provided it meets the rules above. Add a short line for the screen-share segments ("on screen: the pricing spreadsheet at 14:20") so listeners who only hear audio know something visual was shown. Ask the assistant to flag every moment where speakers say "as you can see" or "look at this", since those need a description.

Narrative or scripted show

Where there is a script, use the script rather than the transcript, and have the assistant draft notes from the script's sections. Sources, quotes and archive clips need their own credits list, which should be exact: ask for a table of every clip and its source from the script, and fill gaps by hand.

Handling a two-hour episode without losing the middle

Long transcripts can exceed what an assistant handles well in one go, and even when they fit, quality at the middle of a very long input can sag. Split a two-hour episode into halves at a natural break and run the chapters prompt on each, with an instruction to continue the numbering. Then paste the two chapter lists together and ask one more prompt: "Merge these lists, remove duplicates at the seam, and check that no chapter is under two minutes unless it is the intro or outro." Check the seam by hand: chapters near the cut are where duplicates and gaps appear. In an illustrative 110-minute episode, the split produced two chapters both called "Pricing" at 54:30 and 55:10; a human merged them into one.

A filled-in set of notes and what a reviewer would change

For an illustrative 40-minute interview with the owner of a small print shop, the assistant's first draft contained: summary of two sentences (fine), seven chapters (one off by 40 seconds), a guest line ("owner of a leading print firm") and a list of four links. The reviewer made five edits: changed "a leading print firm" to the guest's own wording, corrected the chapter time at 22:10 to 22:50, replaced one link that pointed to a similarly named supplier, deleted a figure ("saves 60 per cent on waste") that the guest had actually given as "a lot less waste", and added the sponsor line that the template does not carry. Five edits, twelve minutes, and the version that went out had no error a listener could find. The unedited draft had four.

Plan for every first draft to need several corrections; in this illustration it was four errors across seven chapters and four links. Do not trust a streak of clean drafts either: it usually means you are checking less, not that the assistant is improving.

Show notes that also work for search

Notes are text, and text can be found. Two additions help without extra drafting. First, put the guest's name and the topic in the title and first sentence, since people search for both. Second, if your host supports a transcript or a page for each episode, publish the corrected transcript, not the raw one; the glossary pass you ran in the first step also protects your search presence, because a company name spelt three ways is found under none of them. Do not stuff keywords; one clear summary does more than a keyword list, and listeners can tell the difference.

Counting what the template saves over a season

Take a 24-episode season at about 70 minutes of notes each: 28 hours. With the template and checks, 38 minutes each: about 15 hours. The saving is roughly 13 hours across the season, close to two working days. Setting up the template, glossary and checklist takes around three hours, so the payback arrives in the first season. The bigger gain is consistency: every episode has the same shape, so listeners and sponsors know where to find things, and a new producer can pick up the show with a one-page guide.

Errors that survive a casual check, with how they show up

• The merged chapter. Two topics under one timestamp. A listener taps the time and the topic they wanted is two minutes in. Fix: one topic per chapter.
• The plausible link. A link that goes to a real page about something else. Fix: click every URL and read the page title.
• The confident name. A guest's surname spelt as the assistant guessed it. Fix: copy names from the guest form, not the transcript.
• The promoted claim. "Helped many clients" becomes "helped hundreds". Fix: the claims check above.
• Stale housekeeping. A sponsor line or discount code carried over from last week. Fix: put sponsor details in a per-episode checklist, not the reusable template.

Reusing the notes without extra work

Once the notes are correct, they are the best raw material for other posts: a newsletter paragraph, three social posts, a written tutorial outline and clip captions. Ask the assistant to write them from the verified notes, not from the transcript, so the checking carries through. For clips with captions, AI for podcast clips shows the cutting side; what AI can edit in a podcast is useful before you promise anything about editing in your notes. A sponsor reading the notes is also a reader, so if you sell ads, keep sponsor mentions accurate and consistent with what was recorded; the sponsorship media kit tutorial covers what sponsors want to see.

Setting up the template for a team

Make the system boring and repeatable. Put the template, the glossary, the prompts and a pre-publish checklist in one shared folder. On a business plan of your assistant, store them as a project so everyone uses the same instructions. The checklist is five lines: names checked against guest form, every link clicked, every timestamp listened to, numbers confirmed, flags block deleted. Tick them in the episode's task or in a shared sheet. If you want help designing the full episode workflow, from guest booking to published notes, that is the kind of thing my AI implementation consultation covers.

Show-notes questions producers ask

Can AI write timestamps from a transcript?

Yes, if the transcript has time codes, but it can drift or merge two topics under one time. Always click each timestamp in the audio and confirm the topic starts within a few seconds. Treat the AI chapter list as a draft you verify, never as a finished list.

How long should show notes be?

Long enough to help a listener decide and find things: a two or three sentence summary, five to ten chapters, guest details and every link mentioned. Those parts usually add up to a few hundred words. Length matters less than whether each link and timestamp is right.

Should I tell listeners the notes were AI-drafted?

There is usually no need to label routine drafting if a person checked it, but do not let AI claim things the guest did not say. If the notes contain a summary the guest has not seen, consider sending it to them first. Your own policy and any sponsor terms may require more.

Will podcast apps show my timestamps as chapters?

Spotify does when the description lists at least three timestamps, the first at 00:00, each on its own line in time order with a plain-text title. YouTube reads description timestamps as video chapters under similar rules, with each chapter at least 10 seconds long. Other apps have their own rules, so check each one's creator help pages.

Further reads

• Podcast Audio to Video With AI: Waveforms, Captions and Clips — Waveforms, captions and clips for audio-only shows.
• Email Newsletter QA Checklist Before You Hit Send — A final check routine you can borrow for show notes.
• Podcast Guest Outreach With AI: Research, Pitch and Follow-Up — Research guests with AI, send a three-email sequence with real personalisation, track replies, and stay on the right side of email rules.
• AI That Creates Podcasts: Can It Replace a Production Company? — AI can turn documents into a podcast episode in minutes. Here is how to test whether it is good enough to publish for clients, and when to hire humans.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Spotify for Creators help article on episode chapters and YouTube Help on video chapters, both read on 4 November 2026; the template and checks are working method.