AI UGC
Workflow

Before you touch the canvas, decide what a person will give the workflow and what they should get back. Keep it brutally simple: Facebook URL in, ten new ads out. Or website URL in, talking-head avatar out.

This example is deliberately a little complicated. I want to show you the power of a proper workflow and teach the nodes that matter most, including Text Box, Upload Images, Generate Text, Generate Image, Generate Video, Split Text and List Selector. You might not need this exact machine today. If you build it step by step, you will understand the pieces needed to build the machine you do need.

For this build, we are putting in a website URL and a clean product image. We want one believable character, three scenes, a 30-second script, three A-roll clips, two B-roll shots and a review edit out the other side.

Inputs also decide how much control the user gets. We will infer the avatar from the website, but imagine you are building an app that lets somebody choose hair colour and gender. Add a Text Box for each preference, connect both to the Avatar Prompter and tell the LLM to treat those inputs as priority before it executes the remaining instructions.

If you want to inspect the finished routing while you build, open either workflow below. V1 is the original course build. V2 is the updated version I kept improving as I tested it.

Open a blank canvas. In Pletor, right-click anywhere or use Add Node. Add a Text Box for the website URL and an Upload Product Images node for the clean product reference.

Label every node as you add it. When you select a node, its Inputs panel shows exactly what is feeding it. Use that panel instead of trying to trace a wire across the whole canvas.

Paste the website URL as plain text using Ctrl + Shift + V on Windows or Command + Shift + V on Mac. I learned this after a men's Rogaine page somehow produced a woman. I never got a satisfying technical explanation, but pasting the link as plain text fixed the problem.

You can upload more than one product image when the model needs extra angles or label detail. You can also add the preference Text Boxes we discussed in Step 1. For convenience, select the input nodes and group them as Inputs. Grouping is optional, but it makes the board much easier to move and inspect.

The first generated part of the workflow is the person who will carry the advert. Drag from Input Website URL into Generate Text, turn on web search and name the node Avatar Prompter.

I recommend GPT-5.5 for this job. That is my recommendation, not a rule. If you want, duplicate the Avatar Prompter node and split-test the available text models using exactly the same inputs and prompt. I do this a lot. Keep the model that reads the page and follows the output format most reliably.

For now, do not connect the product image. First approve the identity, room, clothing and rough phone-camera look. Then add the real product in the next step. Separating those decisions gives you much more control.

I already explain how to prompt in my prompting ebook, and that guide is comprehensive, so I am not going to waste your time repeating it here. For now, use the copy-and-paste Website to Avatar prompt below and pay attention to what is being automated.

If you added inputs such as Hair Colour or Gender, connect them to Avatar Prompter too. The prompt tells the LLM to treat those choices as priority instead of guessing from the website.

Now we give the approved avatar the real product. The difficult part is scale. A clean packshot does not tell the image model whether the bottle is six inches tall or the size of a suitcase.

You can ask the LLM to read the website, interpret the real product size and use case, then write that detail into the image instruction. It is slightly overkill, but it is more precise and it demonstrates why a text model can be useful between an input and an image model.

Create a GPT-5.5 Generate Text node called Product Compositing Prompter, enable web search and feed it Input Website URL, Avatar 1 and Product Image. Its output then joins Avatar 1 and Product Image inside a Nano Banana Pro image node called Starting Frame 1.

I use Nano Banana Pro here because sending a GPT Image 2 result back through GPT Image 2 can add another layer of texture. Nano Banana Pro did a better job of preserving the person while changing only the product.

Starting Frame 1 locked the person and product. Now we change the room, pose and expression without changing the identity, outfit or rough phone-camera feel.

I am doing this for two reasons. First, I want to show you how to keep one character consistent while changing a scene. Second, changing locations is a common client request because a new background helps reset attention and improve retention on social media.

Add a Text Box called Scene 2 Location Prompt and duplicate the Starting Frame 1 image node. Name the new output Starting Frame 2. Feed it Starting Frame 1, the original Product Image and Prompt 03A: Scene 2 Location Prompt.

The previous frame protects the person. The original packshot restores clean product detail that may already be too small or blurry in the first scene.

You could connect Starting Frame 2 directly to Starting Frame 3 and it would work. I also provide Starting Frame 1 because more approved context gives the image model a clearer picture of what must remain fixed.

The two earlier frames show the same person and product from different poses and rooms. That makes identity consistency easier and also shows the model two locations it should not repeat.

Add a Text Box called Scene 3 Location Prompt, then duplicate Starting Frame 2 and name the new output Starting Frame 3. Connect Starting Frames 1 and 2, the original Product Image and Prompt 03B: Scene 3 Location Prompt.

Wait until Scene 2 is approved before running this branch. A bad frame is still context. It simply teaches the model the wrong lesson.

This is where it can get slippery, so pay attention.

Imagine you want to take one website URL and create ten Facebook ads. If you run ten separate prompts, you can get the same angle twice. If you ask for ten substantially different angles in one prompt, the model can see the whole set and avoid repeating itself. But then you need a reliable way to extract each result. That is what Split Text and List Selector are for.

In this build we need one coherent 30-second script, split into three parts:

It is the same script written in one pass, then split. That keeps the argument, facts, tone and accent congruent instead of asking three separate writers to invent three adverts.

Create one GPT-5.5 Script Generator, enable web search and feed it the website, product image and all three starting frames. Prompt 04: Three-Part Script Generator uses the caret symbol ^ as a delimiter because that character is unlikely to appear naturally in the script.

Video is where the credits start disappearing, so prove one branch before you release all three. Connect Video Prompt 1 and Starting Frame 1 to a Generate Video node called A-Roll 1. Use First frame mode, 9:16 and enough time for the spoken section. Ten seconds is the safe starting point.

I tested Google Omni and Seedance 2.0 against the same frames and prompts. Omni was faster, cheaper and often more candid. It can skip a word when the script is too long, and teeth can occasionally look odd. Seedance was slower and more expensive, but it sometimes followed the dialogue more accurately. Test the models on the same inputs and keep the output you prefer.

Once A-Roll 1 works repeatedly, duplicate the configured node twice. Remove the copied inputs before reconnecting them. Then pair Video Prompt 2 with Starting Frame 2 and Video Prompt 3 with Starting Frame 3.

A-roll carries the message. B-roll gives you somewhere useful to cut when the viewer has been staring at the same talking head for too long. The important word is useful. Two slightly different product spins are not two ideas.

Create a GPT-5.5 B-Roll Prompt Generator, enable web search and connect Input Website URL and Product Image. You can add an approved frame for room and camera context, but it is optional.

The copy-and-paste prompt below asks for two different shots. It uses the same caret, Split Text and numbered List Selector pattern you just learned. Read the prompt first, then compare it with the generated example.

Start with the cleanest packshot you can find. The course test used a fairly artificial image and both models occasionally damaged the label. That is a model limitation, not a routing problem. Compare models using the same prompt and product before deciding which one works best.

This step is optional. I would not recommend using Pletor's Merge Videos for the final advert. It is useful for checking the three A-roll clips as one performance, but the best results come from downloading the clips and cutting them tightly in a proper editor.

If you want a quick review copy, add Merge Videos and connect A-Roll 1, 2 and 3 in order. Start with clean cuts. I tested fades and slides during the recording, and the transition created a black frame that advertised the edit instead of hiding it. A tiny zoom can work. A bad transition is worse than no transition.

You can connect Merge Videos to Zapcap Subtitles to see a rough captioned version. Keep the words around the bottom-middle of the frame, but not so low that TikTok or Instagram covers them. The right edge is dangerous too because that is where the platform buttons live.

For the finished edit, place the B-roll deliberately and control every cut. If you want a faster caption pass, I use Submagic. It is built for short-form captions and finishing, and I prefer it to forcing the final edit through Pletor. CapCut is another option if you already use it. The Google Drive output failed during my recording, so I would not build the process around that connection until you have tested it yourself.

Until now, you have proved that the Rogaine example works. The final test is brutally simple: replace the website URL and product image without touching the routing.

I swapped in a Boots page for Clinique moisturiser and a clean product image. The same workflow chose a plausible female customer, created three rooms, kept the moisturiser in her hands, wrote a new 30-second script, generated A-roll and B-roll, then produced a merged version with captions.

That does not mean you should hit Run and walk away. Release the cheaper stages first. Read the customer choice, approve the avatar and compare all three starting frames. Then test one A-roll and one B-roll branch before you spend credits on everything else. A reusable workflow still needs quality control. It simply gives you one clear place to fix the problem.

Once a second product can travel through the same connections, you no longer have an elaborate one-product demo. You have a machine you can adapt for more products, more hooks, different placements or an eventual software interface.