Create an AI Instagram ad from scratch using MCPs and Claude

This is episode two of Let's cook. Last week we covered the foundation, skills and the first MCP. This week we built something real with it: a finished Instagram ad for an e-commerce product, made almost entirely inside a Claude chat. Four MCPs do the work. Shopify brings in the product, Higgsfield generates the visuals, Canva handles the assembly, and Meta is where the ad runs.

The ad we made

The episode opens with the finished ad: an Instagram reel for an organic candle. There's a living-room scene with a slow camera zoom and the flame flickering, then a second scene of the candle held in hand with the sage, citrus and rosemary it's made of. It looks like something a small studio would charge real money for.

"95% done with Claude and Higgsfield, and a little bit of manual work in Canva."

The store is a demo we set up on Shopify for the episode: a small candle brand. Nobody photographed a product, built a set or edited footage. The rest of the video walks through how the ad was made, connector by connector.

Chapters

The MCP stack

In episode one, connecting Claude to anything meant editing a JSON config file on your desktop. You don't need any of that anymore. You go to Customize, then Connectors, then Add custom connector, paste a URL, and you're done. We wire up three connectors this way in the episode (we showed Meta's last week, but it changed since, so it's worth covering again):

• Meta Ads, now official. Last week we said Meta had no official MCP. A week later they shipped one, and they're rolling it out incrementally. You add the connector, click Connect, log in with Facebook, and Claude has access to your ad account. That access includes write: Claude can answer "which campaigns are active?" in plain language, and it can also create campaigns and pause creatives. The slash-command skills from episode one run on top of it.
• Shopify, your catalog in the chat. Every Shopify store exposes an MCP at yourstore.myshopify.com/api/mcp. Swap in your store name, paste it as a custom connector, and Claude can browse the catalog and product details. For e-commerce this might be the most useful connector of the episode, because the product data (name, price, description, ingredients) flows into every downstream step without anyone retyping it.
• Higgsfield for generation. Higgsfield aggregates the latest image and video models, third-party plus their own, and picks the best one per job. Think of it as Creative Cloud for AI models. Their MCP page gives you the connector URL, and you need an account with some credits, same as any generation platform. A useful detail: everything Claude generates through the MCP also lands in your Higgsfield library, so you can keep iterating there manually. (We're not affiliates, we just like it.)
• Canva, promising but not there yet. The Canva MCP connects and can surface template galleries in the chat, but it can't finish a layout yet; more on that further down.

Put together, that's an end-to-end pipeline: import the product, generate the creatives, rework and tag them, then push them straight to Meta campaigns. It's the same connector pattern Adside builds on for its own creative-tool integrations.

Generating product shots with AI

With Shopify connected, we ask Claude to pull up the candles. Claude inspects the MCP's tools, finds the product and lists the price. Then comes the one-line request: "Can you generate four product shots for it?" Claude answers that it will set this up through Higgsfield, sends everything over automatically, and comes back with visuals.

The interesting part is what Claude did between those two messages. It had the product details from the store, an Italian garden candle made with orange, sage and rosemary, so it wrote its own scene briefs: an Italian garden terrace at golden hour, herbs in frame, and four different vibes instead of four crops of the same shot. One herb came out as basil instead of what's on the label. AI generation still misses details like that, and a few retries usually get you close enough. If you've ever staged e-commerce product photography (the context, the light, the setup), you know that's the painful part, and here it took one sentence. It's the same approach behind Adside's static ad generator: product data in, on-brand shots out.

So what did all this experimentation cost?

"It's 137 credits, around $5.50. An image is like 10 cents, and a video is between 50 cents and a dollar depending on the model."

Even if a single finished video cost five or ten bucks, that's a rounding error next to the production time it replaces. We've put real numbers on that comparison in the UGC vs. static vs. video guide.

Bringing images to life

We're upfront about what we didn't build: an AI UGC video with a synthetic person talking to camera. That format probably won't age well, and it doesn't feel right to us. Where AI does shine is bringing life to imagery: cinemagraphs, subtle motion, the kind of ad you'd otherwise make in After Effects.

On any generated image in Higgsfield you can click Animate. It passes the image as the reference automatically, and you add a motion prompt. The whole motion prompt was "subtle camera zoom, slow flame flickering", and that was enough to turn the still into a living scene. You don't even have to write the prompt yourself; ask Claude for "a prompt to bring subtle motion" and it writes one. This image-to-motion step is what Adside's video ad generator automates across a whole batch of creatives.

One friction point came up here. After a lot of generation iterations, a chat thread becomes a bad workspace; you end up scrolling forever to find the asset you liked three attempts ago. That gap is part of what we're building Adside to fix: the same MCP-and-AI workflow, with an interface built for creative iteration.

Assembling the ad in Canva

The final ad has two scenes: the living room with the zoom, and the candle held in hand with its ingredients around it. We generated a few tests, picked the keepers, and built a 16:9 video. Then we told the AI to resize it to story, and it did.

"That's the type of stuff that would take two days for a team."

The last step, an Instagram reel with text on top, is where the chain breaks today. The Canva MCP connected fine and even pulled up a gallery of templates inside Claude, but it can't complete the layout work yet. So the finish was manual: open Canva, search "candle e-com template," pick one, and drag the generated videos in as backgrounds. That took a few minutes, and it was the only part of the whole ad a human hand touched. The result can be uploaded as a paid ad or run as an organic reel.

What this means for creative production

The bigger point of the episode is about where the bottleneck in ad creative sits. The slow, expensive part was always production: the shoot, the edit, the resize, the handoffs. With product data flowing in from the store and generation models a connector away, a one-person brand produced studio-grade creative in an afternoon, for about the price of a coffee. The pieces that stayed manual (Canva assembly, picking winners out of a long chat thread) are interface problems rather than capability problems, and interfaces tend to get fixed fast. If the per-asset economics interest you, the cost and production-time math across formats is in UGC, static or video: what to run, when, and why.

Links from the episode

Everything we mention in the video, in one place:

• Open-source ad-ops skills library: github.com/AgenticAdvertising/ad-ops-skills
• Meta Ads MCP: mcp.facebook.com/ads · setup guide
• Higgsfield MCP: higgsfield.ai/mcp · connector URL mcp.higgsfield.ai/mcp
• Canva MCP: mcp.canva.com/mcp · setup docs
• Episode 1: Get started with Claude Skills for marketing
• Adside: try it free
• Watch on YouTube: Create an AI Instagram ad from scratch using MCPs and Claude

Quotes are lightly edited from the episode's transcript for readability.