Teardown No. 104 · Short-form video

5.4M Views From Three Slideshows: How Pedro Pérez taught an OpenAI Dot named Joe to research and rebuild TikTok formats for his apps

Pedro Pérez markets his apps with organic TikTok slideshows, and an OpenAI Dot he calls Joe now handles much of the repetitive work. Joe researches TikTok from his own account on his own cloud computer, saves viral references and explains why they work. He then rebuilds a chosen slideshow around Pedro's app using real app screenshots and makes variations of the winners. Pedro reviews every slide before anything is published and schedules the finished posts in Postiz. He reports that three posts made with this system pulled 2.3M, 2.0M and 1.1M views in a few days, more than 5.4M combined.

Pedro Pérez · PerezHatesAIVerified Nov 11, 2026Markdown copy

Try this step

Create one Dot with one job. He named it Joe and sent a job description as the first message: research viral TikTok content and formats, analyze why each one works, and construct ready-to-produce content. The prompt says to prefer slideshows over video unless he asks for video.

The problem

Pedro’s method was already working: find formats that are performing, understand why, adapt them to his app, make new versions and test them at scale. The research was the slow part. He describes spending the first hour of the day researching TikTok, plus building every slideshow and variation himself. Instead of asking AI to randomly generate “viral ideas,” he wanted one agent to learn the system he already uses.

The blueprint (as he describes it)

  1. Create one Dot with one job. He named it Joe and sent a job description as the first message: research viral TikTok content and formats, analyze why each one works, and construct ready-to-produce content. The prompt says to prefer slideshows over video unless he asks for video.
  2. Train it by correcting it. Most of the training happens during real work. If Joe analyzes something incorrectly or ignores the visual composition, Pedro tells him. When Joe saved a post with no app in it, Pedro explained why it didn’t belong instead of just deleting it.
  3. Give it its own TikTok account. Pedro created an account for Joe, logged in on Joe’s computer and handed control back. Joe’s computer is separate from his, so the work continues after he closes his laptop. A recurring instruction can have Joe find 20 new viral app posts every morning at 8:00 AM.
  4. Organize the research. Joe sorts references however the project needs: by niche, format, views, hook style, slideshow vs video or how easy the concept is to reproduce.
  5. Replicate a winner with a master prompt. Pedro sends the post he likes (“I liked this video. Let’s replicate it 1:1, adapting it to my app.”). A long operating prompt, kept as a separate file, tells Joe what to preserve and what to change. Joe pulls real app screenshots from a desktop folder, and Pedro uses Nano Banana Pro for the images. The finished files land in a folder for review.
  6. Turn one format into several posts. Joe keeps the structure and changes the hook, scenario, story, emotional angle, wording and app integration. Pedro keeps the strongest variations.
  7. Review, then schedule. Pedro reviews every creative, uploads the finished assets to Postiz and spreads them across accounts and days, checking the schedule before anything goes live.

Stated results

  • Three posts created with the system: 2.3M, 2.0M and 1.1M views over a few days, “more than 5.4M views combined.”
  • He says the account “started moving very quickly.” No follower, download or revenue figures are given.

Caveats

  • The view counts are his own statement. The app and the TikTok account are not named.
  • The article ends by recommending Postiz for publishing, and Postiz published its own version of the playbook the same day, crediting him. Neither page states whether there is a commercial relationship.
  • He sells a Spanish-language Skool community and 1:1 consulting calls, and offers the full master prompt to people who follow him and comment “SLIDESHOW.”
  • He keeps a human review step: “I wouldn’t let Joe generate something and immediately post it on autopilot without me looking at it first.”
  • Handles: X @PerezHatesAI, author of the primary article (verified 2026-11-11). No Bluesky, Threads, Instagram or TikTok account verified.

FAQ

How does Pedro Pérez use an AI agent for TikTok marketing?

An OpenAI Dot named Joe researches TikTok from its own account, explains why viral posts work, and rebuilds chosen slideshows around his app with real screenshots. Pedro reviews everything and schedules the posts in Postiz.

What results did he report?

Three posts made with the system pulled 2.3M, 2.0M and 1.1M views in a few days, more than 5.4M combined, by his own account.

Does the agent post on its own?

No. Pedro reviews every slide before anything is published and checks the schedule in Postiz. He says AI can still produce wrong text, odd compositions or image mistakes.

How did he train the agent?

He started with a job-description prompt, then corrected Joe during real work, one mistake at a time, and added a long master prompt for replicating formats.

Sources

Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.