Automated TikTok Carousels: 87 Posts in 21 Days for About $21

The problem

Preplo is pre-launch, and he calls TikTok its “primary discovery channel.” He writes that “TikTok’s algorithm rewards consistency,” meaning daily consistency and multiple times daily. He says he can’t design and doesn’t have time to make three posts a day by hand. “For a solo founder, that’s either unsustainable or automated.” He chose automated.

The blueprint (as he describes it)

Run on a schedule. The pipeline runs on cron jobs at 12:00 PM, 4:00 PM and 8:00 PM. Each run generates a complete 6-slide carousel, applies text overlays, uploads it to TikTok’s drafts inbox and pings him on Telegram with a preview.
Generate images with a fixed prompt formula. It took about 40 failed generations before the prompts worked. He credits Oliver Henry, an indie dev getting 500K+ views on AI-generated home renovation carousels, for the prompt approach. Every prompt:
- starts with “iPhone photo of…”;
- uses an obsessively specific kitchen description, copy-pasted across all 6 slides for visual consistency;
- adds lived-in details like a coffee mug or scattered herb stems;
- avoids photography jargon such as lens or depth-of-field terms;
- asks to keep the top 35% of the image relatively empty for text;
- never asks the model to render text.
Tell a story across 6 slides. The image sequence is a hook (phone with cooking video), a video close-up, a messy cooking struggle, discovering the app, the extracted recipe, and the finished dish.
Add text in code, not in the image. A Python script of about 200 lines puts white Poppins Bold text with a black stroke in the upper-third safe zone. It auto-scales font size and processes all 6 slides in under 2 seconds. The text arc runs Hook, Context, Problem, Discovery, Solution, Payoff, at 4 to 8 words per line, and the caption carries the detail. He can change text without regenerating images, which makes A/B tests possible. His first posts had text baked into the AI images. He says the overlay script was 3 hours of coding.
Upload to drafts. The upload is one API call with six image files, a caption and hashtags. Images are converted from WebP to JPEG because TikTok’s API doesn’t accept WebP. The full run takes about 3 to 4 minutes per carousel, mostly image generation latency.
Keep a human checkpoint. Posts go to drafts, not straight to publish. He reviews the images, adds trending audio and publishes. “The AI handles the 95% that’s mechanical. I handle the 5% that requires judgment.”
Pull analytics every morning. At 8:00 AM the system pulls analytics from TikTok’s API, appends a snapshot to a CSV and sends a Telegram briefing. The briefing has total views, top performer, engagement changes and content adaptation notes. He writes, “Automation without measurement is just spam on autopilot.”
Tag every post and kill losers. Each post is tagged with hook_style, recipe_type and cta_style. His framework says “5 data points is enough” to stop producing a dead format. Bold-claim and generic hooks were removed from rotation.

Stated results

87 carousels published over 21 days (522 individual images generated).
142,000+ total views from a zero-follower account.
281 likes and 1 share. He says “low engagement is normal for a new account” and calls views the leading indicator.
About $21 total for image generation.
About 60 seconds of human time per post (review, add music, publish).
Average views by hook style: pain-point ~466, question ~346, curiosity-gap ~297, story ~270, bold-claim ~261, generic ~25. He puts the gap between best and worst at 19x.
Average views by recipe type: international ~830, trending-viral ~782, comfort ~780, dessert ~430.
Week 1 averaged about 400 views. By Week 2 averages stabilized around 500 to 800, and some posts clear 800+.
3 hook styles killed and 2 confirmed winners.
App installs or signups from the views: None stated.

Caveats

This is a first-person Medium post. The carousels promote his own app, Preplo, and include an app discovery slide. The post does not sell a course or kit. It does not say whether he has any relationship with the tools it names.
All of the content is AI-generated, and he says so.
The $21 covers image generation only. Other costs, such as the upload API, are not stated.
He says three hook styles were killed but names only two (bold-claim and generic).
He says he tested too many variables at once in Week 1. With one-variable-at-a-time testing from Day 1, he says he would have found the pain-point winner 4 to 5 days sooner.
He credits Oliver Henry’s prompt approach. Oliver Henry’s handle is not verified and is left out.
This is a scheduled pipeline built on cron jobs and APIs. The post does not describe it as an agent.
Handles: X @StatescuRazvan (verified: linked from the end of his own post). No other handles verified.

FAQ

How much human time does each post take?

He puts it at about 60 seconds per post: review, add music and publish. Everything else runs on cron jobs.

What did the system cost?

He reports about $21 on image generation over 21 days for 87 carousels. Other costs are not stated.

Did the views turn into app users?

The source does not say. No installs or signups are stated, and Preplo is pre-launch.

Which hooks worked best?

Pain-point hooks averaged about 466 views. Generic hooks averaged about 25, a gap he puts at 19x.

Sources

https://medium.com/@razvanst/how-i-automated-3-tiktok-posts-per-day-for-my-recipe-app-with-ai-zero-design-skills-and-21-522f5a081c07 (primary: first-person build write-up by Razvan Statescu on Medium, dated Mar 5, 2026)

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

142K Views: How Razvan Statescu automated three TikTok carousels a day for his recipe app