goulahai.com

Articles · Published · By Goulah

How to Turn a Podcast or Video Into a Blog Post With AI (Without Making Things Up)

I'm a content creator and AI strategist, and I've been creating content for my own brands for about 9 years. A long episode or video already has the hard part done: the ideas, the stories, the expert answers. A written version gives people another way to find it and a page they can skim, bookmark, and share.

The catch is that "paste the transcript into AI and publish what comes out" is exactly the kind of page Google warns against. This guide is a workflow for doing it the right way: the AI does the sorting and drafting, you add what only you can add, and nothing gets published that the episode didn't actually say. It uses Google's official guidance on AI content and Anthropic's prompting docs for Claude. The prompts work in any chat assistant that accepts long pasted text.

1. Know what Google says about AI-written pages

Google doesn't ban AI-assisted content. Its guidance says generative AI can be useful for researching a topic and adding structure to original content. But it also says that using generative AI to create many pages without adding value for users may violate its spam policy on scaled content abuse. Source: Google Search's guidance on using generative AI content on your website (Google Search Central).

The same page says generative models don't retrieve facts, they predict a likely sequence of words, so their output can contain inaccuracies. Google calls it critical to manually fact-check and review all AI-generated content before publishing, and says that review also covers the title, meta description, structured data, and image alt text.

Its spam policies list "using generative AI tools or other similar tools to generate many pages without adding value for users" as an example of scaled content abuse (Spam policies for Google web search). So the goal isn't a transcript rewritten by a robot. It's one useful post per episode, checked by a person, that adds something.

2. Pick one question the post answers

A 60-minute conversation covers a lot of ground. A blog post that tries to cover all of it turns into a long summary, and Google's helpful-content page lists "mainly summarizing what others have to say without adding much value" as a warning sign (Creating helpful, reliable, people-first content).

Instead, pick the one question from the episode your audience would actually search for, like "how do I price my first freelance project" or "what should a first-time homebuyer ask a lender." Write it down in the reader's words. Everything else in the workflow is built around answering that question well. If the episode answers two good questions, that can be two posts, as long as each one stands on its own.

3. Start with a clean, timestamped transcript

If the transcript has the guest's name wrong, so will your post. I covered how to get a transcript out of YouTube Studio and clean it up in How to Find the Best Short Clips in a Long Video With AI (steps 2 and 3). Fix names, brands, and jargon, keep the timestamps, and add speaker labels so the AI knows who said what.

4. Have the AI pull quotes and an outline first

Don't ask for the finished post in one go. Anthropic's docs say that for long documents (over 20k tokens), you should ask the model to pull word-for-word quotes first, which grounds its answer in the actual text, and that explicitly allowing it to say it doesn't know can drastically reduce false information (Reduce hallucinations, Anthropic). The prompting best practices also say to put long documents near the top of the prompt, above your instructions, and to wrap them in XML-style tags (Claude prompting best practices: long context prompting).

This first prompt gets the quotes, an outline built from them, and a list of gaps:

<transcript>
[paste your cleaned-up transcript here, with timestamps and speaker names]
</transcript>

<episode_info>
Show: [show name]
Guest: [guest name and what they do, spelled correctly]
Reader: [who the blog post is for, and what they already know]
Question the post should answer: [one line, in the words your reader would search]
</episode_info>

Using only the transcript above:

1. Quote the 6 to 10 lines that best answer the question, word for word,
   each inside <quote> tags, with the speaker and the timestamp.
2. Suggest an outline for a blog post that answers the question:
   4 to 7 section headings, and under each one, the numbers of the quotes it uses.
3. List anything a reader would need to know to act on this that the
   episode does NOT cover. Label this list "gaps".

Don't add facts, numbers, or examples that aren't in the transcript.
If the transcript doesn't answer the question, say so instead of guessing.

The "gaps" list matters most. It shows where your reader would have to go search again, which is another warning sign on Google's helpful-content page. Those gaps are where you add value in step 6.

5. Draft it in your voice

Anthropic's docs say examples are one of the most reliable ways to steer an AI's format, tone, and structure. They recommend 3 to 5 examples, wrapped in <example> tags so the model can tell them apart from your instructions, and varied enough that it doesn't copy patterns you didn't intend (source). In the same chat, follow up with:

Here are 3 blog posts I wrote before. Match their tone, sentence length,
and how they use headings. Don't copy their content.

<examples>
<example>[paste past post 1]</example>
<example>[paste past post 2]</example>
<example>[paste past post 3]</example>
</examples>

Now write the post from the outline above.
- Open with 2 or 3 sentences that say what the reader will be able to do after reading.
- Use the section headings from the outline.
- Put the guest's key points in your own words and keep up to 3 short direct
  quotes, credited to the speaker.
- Wherever I need to add my own experience or an example, write
  [ADD YOUR OWN: what's needed here] instead of making one up.
- Don't include links, prices, statistics, or product names that aren't in the transcript.

The [ADD YOUR OWN] placeholders are on purpose. They mark the spots where the post needs your experience, and they stop the AI from filling them with a made-up example that sounds real.

6. Add what only you can add

This is the step that separates a useful post from scaled content. Google's helpful-content page asks whether content provides original information or analysis, goes beyond the obvious, and, if it draws on other sources, avoids simply rewriting them and adds substantial value and originality. Ways to do that with an episode:

7. Check every claim, then check it yourself

Anthropic's hallucination guide suggests having the model verify each claim by finding a supporting quote after it writes, and retract any claim it can't support. Run this on the finished draft:

Review every factual claim in the draft you just wrote.
For each one, find a quote from the transcript that supports it.
If you can't find one, remove the claim and leave empty [] brackets where it was.
Then give me:
1. A list of what you removed.
2. Any place where the draft states something more strongly than the speaker did.

The second question catches a common problem: the guest says "this might help," and the draft says "this works." Then do a human pass:

8. Write an honest title and meta description

Google's review advice applies to metadata too. Its helpful-content page asks whether the title gives a descriptive, helpful summary and avoids exaggerating or being shocking. Ask for options, then pick and edit one yourself:

Suggest 3 title options (under 60 characters) and 3 meta descriptions
(under 155 characters) for this post. Each one must describe what the post
actually covers. No hype words, no promises the post doesn't keep.

9. Say who made it and how

Google's helpful-content page suggests thinking about "Who, How, and Why." Put a byline on the post that leads to your About page. If AI did a meaningful part of the work, Google suggests that disclosures are useful where someone might reasonably wonder "How was this created?", along with some background on how automation was used. A short line works, for example: "Drafted from the episode transcript with an AI assistant, then edited and fact-checked by [your name]." Google also warns against fake authorship, like made-up names or AI-generated headshots presented as real experts.

10. Embed the episode and link it both ways

Put the video on the page so readers can watch the parts they care about. Per YouTube's help page, on a computer, open the video, click Share, click Embed, and copy the HTML code into your page. YouTube also offers a Privacy Enhanced Mode: change the embed domain from www.youtube.com to www.youtube-nocookie.com. Source: Embed videos & playlists (YouTube Help).

Know what to expect from search. Google's video best practices say a blog post that discusses an embedded video isn't a "watch page," because the video complements the text, and that it's fine to have the same video on a watch page and on another page like an article (Video SEO best practices, Google Search Central). In other words, the post should earn its place as a written page, not as a wrapper for the video.

Then link the other way: add the post's link to the episode's YouTube description and show notes. If you haven't written those yet, How to Write Podcast Show Notes, Chapters, and a YouTube Description With AI uses the same transcript and quotes.

Reuse the quotes for short clips

The quotes from step 4 are the strongest lines in the episode, with timestamps, which makes them a head start on Shorts and Reels. The full workflow is in How to Find the Best Short Clips in a Long Video With AI.

Mistakes to avoid

Official sources

Google and YouTube guidance in this post is from the official pages above, accurate as of the published date. They can change, so check the linked pages before you rely on a specific rule.