How I Create Podcast Show Notes with Riverside's MCP and ChatGPT

For every podcast episode I produce, I write two versions of the show notes. The first is long: a story-style summary with quotes from the guest, small anecdotes from the conversation and the guest's bio. That version becomes the blog post that goes on the website with the episode. The second is a condensed version that fits the description limits on YouTube, Spotify and the other platforms.

I've always wanted to have long conversations with an AI model to sculpt and adjust these notes, and the Riverside MCP makes that possible. It connects my AI to my Riverside account, so the model can read the transcript of my edit directly, pull the chapters, and even create clips for me.

Here is the series of prompts I go through for every episode, from the first hooks to promo posts and vertical clips.

Connect Riverside to ChatGPT

Before anything else, your AI needs access to your Riverside account. Riverside now has a plugin in the AI marketplaces, so this takes a minute. I'm using ChatGPT here, but the same idea works with any agent that supports MCP.

Start a new chat, click Plugins under the message box, and search for Riverside. If it doesn't show up in the short list, choose Connect plugins and then Browse all plugins to open the plugin marketplace. Search for Riverside there and click the plus button to add it.

ChatGPT then asks you to connect Riverside to your account. Click Continue to Riverside and log in to authorize it. You only do this once. Because I was already logged in, I went straight back to ChatGPT with Riverside approved.

Find the Correct Episode Edit

The first prompt tells the AI exactly which recording to work on. I have several studios and many edits, so I'm specific: "Let's locate Edit 23 in my project with Bret Fisher, in the Media Makers Studio." I call it Edit 23 because I've done a lot of edits on this one. If you work the same way, my guide on creating multiple edits from one recording explains why, and organizing studios and projects makes it much easier to point the AI at the right place.

For the model, I pick the latest one available and set it to high reasoning. A few moments later it confirms it found the edit. Now Riverside knows which project I'm talking about, and every prompt after this one works with that same edit.

Set the Context and Identify the Hooks

The first real prompt sets the context. It tells the AI that it's an expert podcast script analyst and social media content writer, and asks it to pick up the transcript of this edit in Riverside. Before the MCP, I would copy the transcript from Riverside and paste it into the chat. That's no longer necessary.

I also replace the host and guest placeholders with real names ("Cristi Cotovan (host) and Bret Fisher (guest)"), because I work on several podcasts and I want the AI to use the correct names. Then I ask it to identify three things:

• The overarching topic of the conversation
• The topic we refer to and come back to the most
• A few good hooks: insightful, interesting moments that create curiosity and could open the episode

This step is a check. It tells me whether the AI has understood the discussion. In this case it nailed it: building a sustainable business from educational content, with Udemy as the main example, and choosing a course topic with real demand as the recurring theme. It even noticed I described the episode as "how to win on Udemy," which is exactly the title I had in mind.

It also gave me five hooks, each a continuous excerpt from the guest of about 30 to 60 seconds, with timestamps from the edit and a recommendation for which one to open with. I usually favor the ones closer to 30 seconds.

Draft Narrative Show Notes

I reveal what I want a step at a time. Because the context is already in the conversation, each prompt builds on what we've done before. The next prompt asks for show notes of around 4,000 characters that cover the topics we discussed. I don't want a bullet list. I want them to read like a story, the way you'd talk to a friend. I also add a safeguard against clichés, overemphasis, exaggerations, hyperbole and corporate language.

The model you use matters here. With Claude, I prefer the Sonnet models for prose. It may just be my impression, but Opus feels more technical to me. With ChatGPT, the latest models do well. Experiment and see which one gives you the writing you like best.

What comes back is a first draft. It reads well, but it's not finished. We add more in the next steps.

Add Guest Quotes to the Show Notes

Next, I ask for quotes. I like to weave short quotes from the guest between the paragraphs of the story. This is what expands the notes into a longer piece that works as a blog post, and it gives the reader a better feel for the episode.

The prompt says: "This is good. Now let's identify a few great quotes from the guest that can be interspersed in the show notes in the relevant places." I also ask for each quote to be its own paragraph and for long quotes to be shortened. Long quotes may be fine for a clip, but in the notes they reveal too much of the show.

There's one line I added after a lot of trial and error: "Do not worry about the 4,000 character limit from this point." Without it, the model sometimes sticks to the original limit and starts shortening the notes to make room for the quotes. I don't want that. I want the quotes added on top.

The quotes come back placed between the paragraphs, each attributed to the guest. Sometimes the model also adjusts the surrounding text so a fact moves out of the paragraph and into the quote.

Fact-Check Claims Against the Transcript and Bio

Guests sometimes misremember details during a show, and transcripts can get names, places and company names wrong. The model then treats those mistakes as facts. So before anything goes out, I fact-check.

The prompt asks the AI to take every claim, figure, number, year and location in the show notes and show me the specific quote in the transcript where it came from. I also paste in the guest's bio and tell the AI the bio is authoritative. If the notes and the bio disagree, the bio wins.

The report is thorough. Here it found one naming correction, the name of the guest's podcast, which was transcribed wrong. It also clarified some of the pricing details. Then it gave me a table with each claim, the supporting passage with its timestamp, and a finding. At the end there's a separate check for anything that came from the bio rather than the transcript.

If anything comes back as unsupported, my next prompt asks the AI to correct it. Then I get the show notes back with the quotes, the corrections and the bio at the end. If the bio keeps formatting I don't want, I ask for plain text.

Verify the Guest Quotes

This step is optional, and I sometimes skip it because the fact-check has already done most of the work. But if you want to be extra sure you're not putting words in your guest's mouth, ask for a side-by-side comparison. My prompt says the quotes may be slightly paraphrased or have filler words removed, but I want a parallel comparison of each quote in the show notes and the exact passage it comes from.

In this case, all five quotes matched the transcript word for word. The only changes were typographic. The AI also explains what it did with each quote, for example when it joined two adjacent transcript segments into one quote or dropped a side note from the middle.

Finalize and Review the Show Notes

When no more changes are needed, I ask: "Give me the complete show notes now, with quotes, and the bio at the end, the finished version." This gives me a clean final version that includes anything I adjusted along the way.

I don't just copy this and call it a day. I read everything as it's generated, and I often make changes before moving to the next prompt. Then I send the notes to the guest, and to the client when it's a client's show, so they can confirm everything is accurate. There are usually a few days between this and publication. So far, about 95% of the shows where I've sent this format of show notes have come back with no adjustments.

Create Episode Titles and Thumbnail Text

The AI now has the whole conversation in its context, so it's in a good position to suggest titles. My prompt asks it to consider the show notes, the topic and the discussion and come up with something clever and insightful that summarizes the conversation, under 60 characters, with "with [guest name]" at the end. I also ask for short versions of two or three words to use on the thumbnail.

When I want titles that create more curiosity, I add an optional set of rules I've collected from groups where YouTube and podcast creators share their analytics:

• Keep the title under 60 characters. Past that, it gets cut off.
• A title that doesn't trigger emotion won't get clicked.
• Negative framing tends to beat neutral framing.
• Numbers can hurt titles. That one is debated, but I like to see what variations it produces.

It came back with several options, each with its character count and a thumbnail version, and recommended "A Great Course Can Be a Bad Business, with Bret Fisher." That's a title I really like. The other options, like "Good Courses Don't Always Sell" or "Why Your Course Isn't Selling," take a bolder or more controversial angle, which can work well for titles and thumbnails.

If you make your thumbnails in Riverside, my guide on creating AI thumbnails with Co-Creator picks up from here.

Condense the Notes and Include the Chapters

Once the guest or client has approved the notes and the title, we move on to packaging the episode for publication. The long version becomes the blog post, which has no length limit. For the platforms, I need a shorter version. Spotify's description limit is 4,000 characters and YouTube's is 5,000, so I aim for 4,000.

If I made changes to the notes, I paste in the version I settled on. Otherwise I ask the AI to use the last version. The prompt asks it to compress the notes to fit within 4,000 characters without removing essential parts or much of the bio, to trim details and comparisons, and to cut anything that gives away too much of the episode.

It also has to include the chapters. YouTube needs them in the description, so they count toward the limit. And the chapters must not be shortened or changed.

You don't have to paste the chapters. The AI can pick them up from Riverside through the MCP, with timestamps. If you'd rather paste them yourself, open the menu at the top of the transcript in the Riverside editor and choose Copy chapters list. My markers and chapters guide covers how to set them up in the first place.

The short version keeps some of the quotes, opens with the story and includes the chapters it pulled from Riverside, followed by the bio. Looking at this one, I should have asked for a shorter bio. The notes themselves are now compact, so trimming the bio to the essentials would have freed up some space. That's another prompt I sometimes add.

Generate YouTube Tags

Tags don't carry much weight on YouTube anymore, but since I have the context, I ask for them anyway. The prompt asks for SEO-friendly, descriptive YouTube tags for the episode, with the strongest ones first, separated by commas and under 500 characters in total.

It prioritizes more distinctive keywords that might match searches. Copy them and paste them into YouTube. If you also want help with titles and descriptions inside Riverside, see how I prepare YouTube metadata with Riverside.

Write Social Media Promo Posts

Next, I ask for a social media post in two versions:

• A longer one for Facebook and other platforms without a tight limit
• A short one of about 200 characters for X and other platforms where space is limited

Both include a [YOUTUBE URL] placeholder that I replace with the real link once the video is published. This is the output I tweak the most, because I'm not always happy with it. This time the short version came out well: "Can your course become a business? Bret Fisher joins me to explore how to succeed on Udemy and whether your course idea has an audience willing to buy it."

Select Moments and Create Vertical Clips

The last prompt selects clips. I ask the AI to use the transcript with timings and consider the quotes, the hooks, the episode's topic, and the major insights and lessons. It should select at least 10 short clips I could use for Shorts, TikTok and Reels. Each one must contain a complete idea, quote or insight, with no fragments.

I also ask it to find two or three moments where I ask a question and the guest answers, give each clip a short title, and create them as separate vertical edits in Riverside. I can't always get the question-and-answer ones, but it's worth asking.

Here it selected 12 moments, including three Q&A exchanges, and listed each with its length and main idea: "Give Away Your Best Content," "Check Demand Before You Record," "Can You Beat the Top Courses?" and so on.

Then it creates those edits inside Riverside through the MCP. It also builds the layouts and switches scenes based on who is speaking. If the model reaches capacity partway through, you may need to retry, but it finishes in the end.

Review the Clips, Branding and Scheduling

Back in Riverside, I refresh the project and open the dropdown with the project's edits. The new clips are in the Magic Clips section, each with the title the AI gave it.

Opening one, it's already vertical. I didn't ask for branding, but you can: ask the AI to apply your branding and it adds your captions and logo from your studio's brand kit. I also didn't ask it to remove filler words. I want these raw, and if I need to cut anything I'd rather do it myself. If a clip ends too early, you can reveal more content in the Magic Clip.

The MCP can now also schedule the clips. You can ask it to put one a day into the Riverside content planner, or on whatever schedule you prefer.

The Whole Workflow at a Glance

It's a long workflow, but it all happens in one conversation with full access to the episode. Here's the order I follow:

1. Connect Riverside to your AI and locate the exact edit.
2. Topic and hooks: check that the AI understands the episode.
3. Story-style show notes of around 4,000 characters.
4. Guest quotes between the paragraphs, ignoring the character limit.
5. Fact-check every claim against the transcript and the guest's bio.
6. Verify the quotes side by side (optional).
7. Finished version, reviewed by you and sent to the guest or client.
8. Titles and short thumbnail text.
9. Condensed notes under 4,000 characters, chapters included.
10. YouTube tags, promo posts, and vertical clips created in Riverside.

Connect your favorite AI to Riverside, give it the work, and review what comes back.

Related guides

Riverside Markers and Chapters: A Complete Workflow Guide

Markers and chapters do different jobs in Riverside. Here's how to drop markers while recording, manage them in the editor, shape chapters in the transcript and on the timeline, and export a chapter list to YouTube and Spotify.

Cristi CotovanCristi Cotovan

How I prepare metadata for YouTube quickly using Riverside

I walk through the exact steps I follow so you can replicate the same process and prepare metadata for YouTube without overworking each upload.

Cristi CotovanCristi Cotovan

How to Reveal More Content in a Riverside Magic Clip

Here's how to reveal more content around the magic clip and include the parts that were trimmed by Riverside.

Cristi CotovanCristi Cotovan

Riverside Content Planner: Schedule Social Media Posts Inside Riverside

Learn how to use the new Riverside Content Planner to schedule YouTube and YouTube Shorts posts, generate AI descriptions, and manage all your content from one calendar.

Cristi CotovanCristi Cotovan

How to Personalize AI Responses and Automations in Riverside

Make the Riverside AI and Automations Yours!

Cristi CotovanCristi Cotovan