Custom client builds
Kris Papa, artist manager · Artist management
Kris opens one page and sees what his artist's niche is discussing online, so he can be part of the conversation.
Kris Papa manages recording artists. He sent me a reel about using AI to find the conversations an artist should be part of, then came over to build it. By the end of the afternoon it was pulling real comments, ranking them, and showing them on a private board.
Map the niche
Pull the comments
Rank them
Kris's boardA make-believe version of his private board
What the niche is discussing
Each comment ranked against its own video.
Comments read
705
618 with real words in them
A make-believe version of the board. The real one is private to Kris.
- Client
- Kris Papa, artist manager
- Sector
- Artist management
- Built
- TikTok comment reader, ranked brief, private board per artist
- Status
- Delivered. Kris owns the code and the hosting
The problem
Everyone says know your audience. Nobody has time to read the comments.
Real audience research is sitting in comment sections: what people argue about and what they keep asking. For one artist's niche that is thousands of lines under dozens of videos.
A chatbot will summarize them for you, and it sounds sure of itself whether or not it read a single comment. Kris wanted receipts, with every claim on the page tied to a real comment.
What I built
Code counts. The model judges.
The whole tool runs on that rule. Scraping and counting are plain code, because they have to be exact and cheap. Working out what an argument means goes to the model, because that takes judgment. Swap the two and you pay more for worse answers.
- 01Map the nichesStart from a profile of the artist written by someone who knows them, and work out which communities and search terms are really theirs.
- 02Pull the real commentsSearch TikTok for those terms and collect the public comments under each video. It only reads public comment sections. It never goes behind a login or into DMs.
- 03Boil them downEmoji-only lines are dropped and repeats are counted. Each comment is ranked against the other comments on its own video, so a small video and a viral one compare fairly.
- 04Find the argumentsThe model reads the brief and names the live arguments, the questions that keep coming up, and the opinions that get pushback. Then it drafts angles the artist could speak to.
- 05Put it on a boardEach artist gets a private page behind a password, and Kris opens it like any website. Nothing in the system posts anything. Every idea is a draft for a person to approve.
What changed
The first version ranked the wrong comments.
The first version ranked comments by raw likes. One video had 11.9 million plays, so its comments buried everything from the smaller ones. Scoring each comment against its own video fixed it. The board looked just as finished before the fix as after, so a tidy page told us nothing about whether the numbers were right.
The research also leaned on old press at first, which pointed at a scene the artist had left years ago. Someone who knows the artist beats the press, and the press beats a scraper. A profile written by a person is now the first input.
That night the code moved to Kris's GitHub and the board moved to his hosting account.
Build notes
- TikTok search and comment collection through Apify, billed per result
- The scraper is plain Python with nothing to install
- Four Claude Code commands: niche map, tensions, angles, board
- One self-contained page per artist, password-protected
- Code and hosting handed to the client the same day
- Reports are dated and kept, so the trend over months shows
Scope and limits
Kris is a friend, and this was one working afternoon, not a long engagement. The tool reads TikTok comments only. The part that tracks which fans come back month after month is designed and not built. The board is private, so there is no public link.
Matt Smith, Chicago. matt@visibilitybuddy.com