AI Ad Creatives at Scale Start With a Company Brain, Not a Prompt
By Raphael Guilhem, Founding Growth at Pletor ·
Open ChatGPT, type "create an image to promote my SaaS," and you get something generic and off brand that you would never ship. Raphael Guilhem thinks that is the wrong place to start. In this session he skipped the slides and worked live in Claude, connected to Pletor through MCP, to show how a B2B team can produce on-brand static ads, localized variations and AI video without bouncing between a dozen tools. The trick is to give the model your brand context once, in a "company brain," and then keep asking for more. This recap covers how the brain gets built, how he turns competitor ads into on-brand versions of your own, the workflow he uses to scale variations, the one-take video method behind his own LinkedIn reach, and what it all costs.
Why do AI-generated ads usually look off brand?
Raphael put it bluntly: most people already know how to make an ad. The trouble is that the ones they get out of AI are never good enough. A cold prompt in a chat window has nothing to work with except the words you typed, so it falls back on stock patterns. You can't reproduce the result reliably, and it doesn't look like your company.
His fix is to stop treating every ad as a fresh prompt. Build up the brand context first, so each generation starts from who you are.
What is a company brain and what goes in it?
A company brain is a single place in Pletor where all of your brand context lives. Raphael's demo brain was built for HubSpot, purely as an example (he does not work there). It holds:
| Brain section | What goes in it |
|---|---|
| Identity | Logo, brand kit, colors, typography |
| Guidelines | Brand book, rules for how the brand shows up |
| Knowledge | Brand core, unique selling points, competitors |
| Products | Each product line, what it is, who it's for |
| References | Ads you have already shipped, plus competitor ads you want as inspiration |
The references part is where the value builds up. If you shipped 100 ads in the last few months, you can load them in, and every new generation learns from them. In his words, the more complete the brain is, the more on brand your ads feel.
How do you build the brain?
There are two ways in. In the Pletor interface, you click plus, paste your website URL, answer a few questions, and it builds itself. You can then upload your brand book, guidelines or anything else you have internally.
Raphael does it from Claude. With the Pletor MCP connected, his whole instruction is a plain-language request to create a Pletor brain for the company. Claude builds it from the website, asks a couple of follow-up questions, and you keep iterating by answering them. From the website alone, his HubSpot example pulled the logo, brand kit, colors, typography, brand core and unique selling points.
Other sources work too. Pletor has a Notion connector, you can route Notion through Claude, or you can export a Notion page as a PDF and upload it. Raphael's view is that anything with company context belongs in there, including client calls.
What if you run several brands or product sites?
A brain holds multiple products, each with its own colors, positioning and target buyer. His example was a nutrition brand with a range of products in a single brain. If your products live on separate websites, paste those URLs or that content into the chat and the brain builds out each one.
How do you keep a mascot or unique object accurate?
Visual fidelity goes in the brain too. Raphael said many Pletor customers have mascots. You add the mascot to the brain, then state in your instruction that every ad must include it. With the context stacked in the brain, he can keep his prompts broad, and get more precise only when a specific element matters.
How do you connect Pletor to Claude?
The setup takes a minute. In Pletor, open the MCP section and copy the server URL. In Claude, go to Customize, then Connectors, add a new one and paste the URL. The same pattern works for any MCP-compatible platform.
Once it's connected, Claude can generate images and video directly, and Pletor runs in the background. Every ad Claude creates also lands on a Figma-like board inside Pletor, so you can view, edit and branch variations visually. Raphael said the main benefit is that the whole team works in one workspace instead of switching between tools.
How do you turn competitor ads into on-brand versions of your own?
This was the core of the static ad demo, and it ran in three prompts:
- Scrape competitor ads into the brain. He had Claude pull the last 10 static ads from Salesforce, Attio and Notion and add them to the brain as references. Now you never start from a blank page.
- Recreate them with your brand. Next prompt: recreate the two latest competitor ads using our brand system, our brain and our unique selling points. Claude returned ads using HubSpot's logo, font and selling points while keeping the competitor's angle. He had Claude attach each source ad next to its output so you can compare.
- Iterate on the variations. From there you edit the copy or targeting directly. Swap "London" for "New York," select a batch and set "target the US only, in US English," or change "AI agents for executives" to "AI agents for sales." You can select ads in bulk and change them all in one request.
One output came back in German. The source competitor ad had been localized for a specific market, and the model carried that over. It's a good reminder to check the outputs.
Raphael was clear that the "recreate the last two ads" prompt is a demo of what plain language can do, not a strategy. You don't know which scraped ads actually performed. Before reproducing anything, curate the set, either by hand or with an agent, and decide which creatives deserve to be copied.
Which model should you use for static ads?
For image ads that include text, Raphael recommends GPT Image 2 and nothing else. He called it the best model for ads and text, and the one that best holds a mascot or visual style consistent across ads. Inside Pletor you pick the model per generation. His flow defaulted to Nano Banana, and he switched it to GPT Image 2 for the ad work, at medium quality in the demo.
How do you scale ads with a workflow?
Once you can produce single on-brand ads, Raphael moves to Pletor's workflow builder to make them in volume. He knows workflows put people off, so he described it as having Claude as an assistant that builds the workflow for you:
- Create a new flow and select your brain, so all the branding and context comes along.
- Write a short text brief, for example: make static image ads for my company by reproducing competitor ads.
- Answer the follow-up questions. The assistant then builds the flow on its own.
- Upload your inspiration ads (he used three, and said you could add six).
- Run it in loop mode. One click produced on-brand variations, for example "Grow your revenue with HubSpot" in place of the Attio original.
- Add market localization nodes for Spanish, French, Italian and other languages.
- Change the aspect ratio by connecting the base image to a new generation node set to the ratio you need.
Can Pletor publish the ads and learn from performance?
Partly. Connecting a Meta account and pushing ads live is in beta and not open to everyone yet. The goal is a feedback loop: see which ads perform, then reproduce and scale the winners.
How do you make a multi-angle video from a single take?
This is the format behind Raphael's viral LinkedIn post: a video that looks like a multi-camera studio shoot, made from one take of him filming himself in an ordinary room. With AI you can change the background, the camera angle and even your appearance.
Get the full playbook: Raphael wrote up the whole method in Become Anyone From Any Angle: The One-Take AI Video Guide, a step-by-step written playbook for filming yourself once and turning it into a multi-angle vertical video.
The workflow he showed:
- Decide who you want to become. If you're changing your appearance, choose the new character.
- Build a character sheet. It shows the new character from multiple angles so the AI understands what they look like.
- Pick a background reference. He used an image as background inspiration and told the prompt to remove the person in it.
- Have Claude write the prompt. He has a Claude skill built from a web page he found with the right prompt structure for Seedance. He gave Claude his raw video, the character and the background, and explained the goal in plain language: same scene, new variables. Claude first proposed the multi-camera angle changes. Once he approved them, it wrote prompts timestamped for each part of the video.
- Run it in Pletor with Seedance through the Pletor MCP.
His advice for getting started is to paste the guide into a Claude Code session and tell Claude to follow it with your own video and background. Because he knew the exact process, he got his result in one shot, and he expects the guide will get you close to that too.
Why use Pletor if Seedance is the model doing the work?
Raphael said that for this kind of video nothing matches Seedance right now. Pletor's job is to bring the best image and video models on the market into one workspace, so you can orchestrate everything from Claude in one place. You could call Seedance directly if you prefer. He also doesn't see new models as a competitive threat, because Pletor gets better whenever the models it runs get better.
How do you recreate a winning video ad with a new model or outfit?
His second video example came from a fashion brand. They had a winning ad with one model in a blue outfit, from a shoot he guessed cost around 20k. The question was how to make more versions of that winner without another shoot.
His only input was the original video, a new AI-generated model and a new outfit. He sent the video to Gemini, which he described as the only model that can read a video, to break it down with precise timestamps. That breakdown went to Seedance, which produced the same ad with the new character in the new outfit, keeping the product and the person consistent throughout.
Product and character consistency is the hardest part of AI video, and this is the use case he picked to show it.
What does it cost to produce AI ads this way?
| Asset | Cost Raphael cited |
|---|---|
| Static image ad | A few cents each |
| Video ad | A couple of dollars |
| Seedance video | About 1,000 credits for 30 seconds, roughly $1 per second of video |
| His one-take multi-angle video | About $20 |
Seedance is the best model for this work and also the most expensive, so Raphael uses it for short-form content, ads under 30 seconds. Most of the cost comes from iterations. As he put it, failing usually costs more than succeeding, which is why he recommends following a proven process instead of experimenting from scratch. The $20 video brought in a lot of leads, which is part of how he ended up presenting at this session.
Pletor gives free credits on sign-up. He believes the brain is not part of the free version, so plan for a paid plan if you want the brain and MCP workflow he demoed.
Does this actually work in production?
Pletor has more than 1,000 clients and expects to reach 2,000 soon. Raphael's argument for why AI fits ad creative so well comes down to asymmetry:
- Failing costs almost nothing. A failed static ad costs a few cents. A failed video costs a couple of dollars.
- A winner can be multiplied. Once something wins, you don't stop there. You duplicate and amplify it.
For B2B SaaS, he gave an example. If a static ad with someone's face and a line of text is winning, connect the MCP and ask Claude for 10 variations on the price, 10 on the ICP and so on, at a fraction of the cost of new production. His best clients send out 20 variations as soon as they have a winner. In his view, the teams using AI to ship faster are the ones winning right now.
What about copyright?
Copying competitor ads raises the obvious question. The model in the fashion example was fully AI-generated, and the scaling advice Raphael gave centers on iterating your own winning ads. Treat competitor ads as inspiration for angles and formats, and rebuild them around your own brand and selling points rather than cloning them.
Key Takeaways for GTM Teams
- Build the brain before you generate anything. Website, brand book, products, mascot, past ads and client context. The more complete it is, the less you need to prompt.
- Connect your creative tool to Claude through MCP so ads, variations and video come from one workspace in plain language.
- Use competitor ads as references, not a script. Scrape them into the brain, curate which ones are worth reproducing, then rebuild them with your own brand and USPs.
- Pick the model for the job. GPT Image 2 for static ads with text, Seedance for short-form video, Gemini for breaking an existing video down into timestamps.
- Scale with workflows. Loop mode, localization nodes and aspect ratio changes turn one ad into a full set.
- When something wins, flood it with variations. Failing costs cents and dollars, and a winner can be multiplied almost for free.
Want to Attend the Next Session?
GTM Clawllective is a community of go-to-market operators sharing how they actually use AI, live and hands-on. The live discussion stays off the record, and every recap includes the speaker's demo video so you can watch the build yourself. To catch the next one live, join the community.
About the Speaker
Raphael Guilhem is Founding Growth at Pletor, a platform that lets anyone create product images, video and ads at scale, with the leading image and video models in one workspace. He runs his own creative operations through Claude connected to Pletor's MCP, and his one-take multi-angle AI video format took off on LinkedIn.
Read his playbook, Become Anyone From Any Angle: The One-Take AI Video Guide, and connect with Raphael on LinkedIn.
Watch the session replay: How to Create Ad Creatives at Scale With AI