How I Built a US TikTok Slideshow Factory That Researches Formats and Produces Every Day

Cover image

As many of us have come to realize, distribution has become the main bottleneck of founders today. Looking for content on TikTok is a job in itself and yet some people claim it is easy as hell to do it well by just copying stuff. The truth is it's not that easy

Ideas are cheap now. Hooks are everywhere. Formats can be reverse-engineered. Images can be generated. Captions can be written in seconds.

The hard part is building a system that keeps producing when you are not looking at it.

That is what I wanted to solve: not “how do I make one good slideshow?”, but “how do I build a factory that can research what is working, turn that research into TikTok slideshows, keep each account differentiated, and prepare content every day without me manually babysitting the process?”

The result is a pipeline built around two separate jobs: weekly research and daily production.

The research layer finds promising TikTok formats. The production layer turns those formats into slideshows. The publishing layer prepares the final package for TikTok. And the whole thing is designed around one important principle: the factory must not stop producing just because research does not run that day.

The factory mindset

Most content systems break because they are built like one-off prompts.

You just ask your agent for an asset, and idea, then you go for the outcome and it if work you repeat it the next day. That's not efficient as it might work for a few days but not in the long run or if you wanna scale things up.

I wanted something closer to a production line:

once a week, the system studies what is working;
it stores enough hooks, themes, and formats for the days ahead;
every day, it consumes that inventory and creates a new slideshow;
each slideshow has its own concept, caption, visual direction, and CTA;
the system checks that the output is safe for TikTok before delivery.

This distinction matters. Trend research should not be a daily blocker. It should create inventory. Daily production should keep running from that inventory.

If the weekly research step fails, production should still continue using the last valid research batch. A content factory cannot depend on one fragile step executing perfectly every single day.

The two engines: Trendwatch and Carousel Conveyor

The pipeline is built around two skills.

The first one is Trendwatch. Its job is not to invent content from scratch. Its job is to observe TikTok, find patterns, and extract formats that are already showing signs of working.

The second one is Carousel Conveyor. Its job is to take those findings and turn them into actual TikTok slideshows: slide structure, overlay text, image prompts, CTA selection, caption, and final assets.

I keep those jobs separate because they require different kinds of thinking.

Trendwatch is strategic. It asks: what hooks are getting attention? What formats are reusable? What topics are aligned with the product? What kind of first slide earns the scroll stop? For it to work I give Codex access to "Computer Use" so it can browse the Internet with any browser and check TikTok as a real person would.

Carousel Conveyor is operational. It asks: what should slide 1 say? What should slide 2 show? Where should the overlay text go? Which CTA should be used? Does the final slideshow fit TikTok’s safe zones?

Mixing those two jobs into one prompt makes the system harder to debug. Separating them makes the pipeline easier to control.

How the weekly research works

Once a week, Trendwatch analyzes TikTok using a browser workflow.

The goal is not to scrape private data or pretend we can see metrics that TikTok does not expose, but to use visible signals and format analysis: hooks, topics, visual structures, comments, repetition patterns, and public engagement cues.

A good Trendwatch run gives me enough material for the whole week: a backlog of hooks, angles, and slideshow ideas that the daily factory can consume one by one.

The structure is as follows:

the hook;
the reference format;
the product angle;
why the format might work;
the intended engagement reason;
the suggested slideshow structure;
notes for what should be tested.

That last part is important. Each slideshow should test one main variable. If everything changes at once — hook, topic, visual style, CTA, caption — you learn very little from the result.

How daily production works

Every day, the production layer picks one valid concept from the research inventory and turns it into a TikTok slideshow.

This is where Carousel Conveyor takes over.

The slideshow does not have to be the same length every time. Some concepts need two or three slides. Others need six or seven. The point is not to force a rigid template. The point is to build the minimum number of slides needed to deliver the idea clearly.

Each slide has a job.

Slide 1 has to stop the scroll. The middle slides have to carry the story, comparison, or useful insight. The final slide has to reconnect the viewer to the product without sounding like a hard ad.

That final part is where a lot of AI-generated content fails. It creates content that could exist without the product. If the post would make sense even if the app did not exist, the concept is too generic.

The product has to be central to the payoff.

Captions and overlay text are not the same thing

One of the most useful rules in the pipeline is that captions and slideshow overlay text are handled separately. I need a very specific caption for my posts as it includes my download links and promo text.

The caption is the text that accompanies the TikTok post.

The overlay text is the text placed directly on the slideshow images.

They have different jobs.

The caption can add context, push discussion, or frame the post for TikTok. The overlay text has to guide the viewer through the slides in real time.

For that reason, I do not use caption rules to write the slide text. The slide text follows the slideshow structure created by Carousel Conveyor based on the Trendwatch finding.

This keeps the output cleaner. The post caption can sound natural and platform-native, while the slides themselves stay focused on the visual story.

Images are generated clean, then text is rendered locally

Another important rule: text should not be baked into the generated images.

The generated image should be clean. No fake typography, no distorted words, no duplicated text, no unreadable AI lettering.

The overlay text is added later by the local rendering pipeline.

This gives me control over:

exact wording;
font size;
line wrapping;
TikTok safe margins;
vertical placement;
final export size;
readability.

It also means I can fix a typo without regenerating the image.

For TikTok, this matters a lot. Text that looks fine in a local preview can be hidden by TikTok’s interface once uploaded. So the renderer has to respect safe zones and place the overlay lower or higher depending on the slide type.

The final asset is always rendered as a TikTok-ready vertical image, not as a random generated image that happens to be close to 9:16.

Persona consistency

For accounts that use a recurring human persona, the model cannot be treated as a vague prompt.

If the account is supposed to use a specific young male persona, the system must use that persona. Not “a young person”. Not “a lifestyle creator”. The actual approved reference.

That is now a hard rule in the pipeline.

The persona reference is attached as a real visual reference during generation. The system also checks that the final slideshow did not drift into the wrong character, wrong gender, or wrong visual identity.

This is especially important when scaling across multiple accounts. Each account needs its own identity. TikTok can penalize repeated content, and viewers can feel when an account has no coherent voice.

The goal is not to blast the same slideshow across every account. The goal is to create different slideshows from the same underlying research engine.

Same research layer. Different account outputs.

The role of Codex in the pipeline

In the original inspiration for this workflow, Claude acted as the main brain.

In my version, Codex is the operating center.

Codex reads the research, applies the content rules, generates or coordinates the assets, checks the final package, and hands off the slideshow to the local TikTok delivery flow. On top of that Codex itself generates the images with Image-gpt-2.0, thus I don't need to call external APIs and everything remains native and free.

The local pipeline handles the deterministic parts:

file structure;
asset validation;
image dimensions;
overlay rendering;
duplicate checks;
caption output;
delivery package creation;
run state;
failure reporting.

This split is important.

The AI model is good at judgment: hooks, angles, captions, visual ideas, QA.

The local scripts are good at contracts: exact filenames, dimensions, state files, delivery requirements, and repeatability.

A reliable automation needs both.

Preflights before spending generation

Before the pipeline spends image generation or prepares a post, it runs checks.

The system verifies that the target account is correct, that the persona reference is available, that the CTA library exists, that there is trend material to consume, and that there is no same-day state blocking the run.

This prevents expensive or risky failures.

For example, if the TikTok profile is wrong, the pipeline should stop before rendering. If the persona reference is unreadable, it should stop before generating. If a slideshow has already been prepared for that day, the system should not create a duplicate unless explicitly asked.

The basic rule is simple: validate cheap things before doing expensive things.

Deduplication is part of the product

Because the goal is to run multiple accounts, deduplication is not optional.

The system should avoid:

sending the same slideshow to several accounts;
repeating the same first-slide visual too often;
reusing an identical hook on the same day;
creating near-duplicate packages that TikTok could interpret as spam.

This does not mean every post has to be completely unrelated. It means each output needs a meaningful variation: a different hook, visual framing, story angle, CTA, or product payoff.

The research can be shared. The final slideshow cannot be copy-pasted everywhere.

The final step is not “post whatever was generated”.

The pipeline prepares a TikTok-safe slideshow package and sends it through my local TikTok draft workflow.

That matters because I do not want the system to publish blindly. I want it to prepare the content and place it where it can be reviewed, finalized, or sent through the approved flow.

The package includes the rendered slides and the caption. It also records enough state to know what was generated, when, for which account, and from which concept.

If delivery fails, the run should report the exact failure instead of pretending everything worked.

What this gives me

The main benefit is not that I can generate slideshows faster.

The benefit is that the process compounds.

Every weekly research run improves the backlog. Every daily production run gives me another test. Every output is tied to a concept, a hook, and a format. Over time, the system becomes less like a random content generator and more like a structured experimentation machine.

It gives me:

consistency without daily manual work;
a reusable research base;
account-specific content instead of duplicate spam;
faster testing of hooks and formats;
cleaner separation between strategy and production;
fewer failures caused by missing assets or wrong accounts;
a pipeline that can scale to more accounts.

This is the part I care about most: the automation lets me keep distribution running while I focus on the product.

A prompt can make a post.

A factory can keep producing, keep state, avoid duplicates, recover from failures, and turn content into a repeatable operating system.

That is the real leverage.