Berman's creator vetting prompt sorts on buying intent

Matthew Berman published a one-prompt creator scout that takes a creator's profile, their last 12 posts, transcripts of their top 3 videos and the top 50 comments per post, then scores the creator 1 to 5 on five axes and quotes the evidence for each.

Matthew Berman

@TheMattBerman

X

a 40k creator whose comments ask "where do I get this" beats a 400k creator whose comments are fire emojis

The five axes are audience match, buying intent, on camera behaviour, consistency and red flags. Audience match asks the model to quote 2 comments from people who look like the buyer. Buying intent is a count of comments asking where to buy, the price, or for a link. On camera separates creators who explain why they like a product from creators who just hold it up. Consistency checks whether they posted in the last 14 days and compares median views against follower count. Red flags include giveaway bait comments, 3 or more sponsored posts in 30 days, and engagement that is way off the rest of the list.

The output is a KEEP, MAYBE or CUT call with one line of reasoning, plus a two line DM that has to name one specific video. Berman explicitly bans the phrase "love your content" in the DM.

The two instructions worth stealing

Two lines in the prompt do most of the structural work and transfer to any scoring task. The first is the missing data rule, which tells the model to write MISSING rather than produce a number it does not have. The second is the sort key. Berman says to run the prompt on every handle in the list and sort by buying intent, not followers.

His argument for that ordering is the whole pitch. Follower count tells you reach, the comments tell you if anyone buys, and the intent count does half the job of ranking a list.

What this is not

This is a prompt and a feeding recipe, not a measured result. Berman suggests pulling the posts, transcripts and comments with the ScrapeCreators MCP or an Apify TikTok or Instagram actor, and he gives a worked example of the brand slot filled in for a protein bar company selling to women who lift. He does not report an accuracy figure, a cost per creator, or a comparison against any other vetting method, and he does not name the model he runs it on.

For people building on it, the scoring rubric only works if the scraper actually returns 50 comments per post and three transcripts. Feed it a thinner pull and the MISSING token is doing the work instead of the scores, which is the correct failure mode but not a useful one.