How to Train an AI Persona on Your YouTube Videos

Most creators assume training an AI on their channel means writing a long description of themselves. It does not. The persona learns from what you already said on camera, and the only writing you do is a few lines of guidelines at the end.

Step 1: Make sure your videos have captions

The persona learns from transcripts, and transcripts come from captions. YouTube auto-generates captions for most spoken-word videos, so for the majority of channels this is already done. The exceptions are music videos, videos with no speech, and videos where you turned captions off.

If a video has no captions, it contributes nothing. Open YouTube Studio, check the Subtitles tab on your most important videos, and turn on auto-captions where they are missing. Do this before you paste the link, not after.

Step 2: Paste your channel URL

Any channel format works: the handle URL, the old channel ID URL, or a custom name URL. Staraze lists your uploads, fetches the transcript of each captioned video, and records the title, publish date and word count alongside it. A training run covers your most recent videos, roughly fifty transcripts, which is more than enough text to establish how you talk and what you know.

This is the one step that cannot be rushed. A long back catalogue can take a few minutes to fetch. Leave the tab open.

Step 3: Add the sources YouTube does not give you

Instagram, TikTok and Facebook block scrapers, so a link to those profiles yields little more than the bio line. Treat them as a signal of who you are, not a source of knowledge. The sources that genuinely add depth are:

• Your website or blog. Written articles are dense, edited versions of what you think, and they fill in topics you never made a video about.
• A PDF. A course workbook, an ebook, a recipe collection, a training programme. One PDF can outweigh twenty short videos.
• Long-form posts. A newsletter archive or a text thread you are proud of.
• Episodes, if you have a show. Podcast transcripts are the densest source of all; see AI personas for podcasters.

Step 4: Check what the analyser detected, then correct it

After the scrape, Staraze reads the material and proposes a personality profile: tone, recurring topics, catchphrases, how you open and close an answer. Most of it will be right. Some of it will be a caricature, because a persona analyser weights whatever you say most often, and most creators say their intro line more than anything else.

Edit three things specifically. First, the topics you want it to lean on, which should match what fans actually ask you. Second, the topics it must refuse, such as medical dosing or anything you are not qualified to answer. Third, the sign-off. A persona that ends every message with your video outro gets old in two replies.

Step 5: Test it like a sceptical fan

Ask it the three questions you get most in comments. Then ask it something you have never covered. A well-grounded persona answers the first three in your voice with your actual positions, and admits the fourth is outside what it knows. If it confidently invents an opinion on the fourth, tighten the guidelines to say so and retrain after adding content on that topic.

Retraining is the fix for most quality complaints. Every time you upload a batch of new videos or publish a new guide, run it again. The persona only knows what was there the last time it looked.

Step 6: Verify, price and publish

Before a persona goes public, Staraze verifies the creator behind it. That identity check exists so nobody can publish a persona of you without you. You also set at least one message pack, for example a price and the number of messages it includes, because a persona cannot go live without a way for fans to pay for it. Then share the page link. Visitors can try three free messages before they are asked to sign up, so the first impression does the selling.

If you want the business side first, read whether an AI version of yourself can make money. If you want to understand the mechanics, how creator AI chatbots work covers retrieval in plain English.