The Agent That Learned How a Client Wanted Their Posts to Look
The Agent That Learned How a Client Wanted Their Posts to Look
The client did not come to us because they needed more content ideas.
They had ideas. They had offers, product photos, campaign themes, and a calendar full of things to publish. The problem started after the idea was approved, when someone had to turn it into an actual post design.
Some designs came from templates. Some came from freelancers. Some were generated with AI image tools. And many of them came back with the same comment from the client: “It is close, but it is not us.”
The layout was clean. The colors were acceptable. The image looked polished. But the post still felt wrong. The product was too small. The background was too loud. The typography did not carry the brand. The spacing felt like a stock template wearing the company’s logo.
The Problem Was Visual Taste, Not Image Quality
At first, everyone tried to solve it with better prompts.
They described the brand colors, uploaded references, asked for “premium but friendly,” “modern but not cold,” and “bold but not messy.” The AI could generate beautiful images, but beauty was not the problem.
The client had a very specific visual instinct. Certain shadows felt cheap. Certain product angles felt weak. Some posts needed a lot of negative space, while others needed dense offer treatment. A background that worked for one campaign looked completely wrong for another.
So instead of building a generic image generator, we built a social media design agent around the client’s real visual memory.
Teaching the Agent What “On Brand” Looked Like
We collected approved posts, rejected designs, product photos, campaign boards, logo usage rules, font choices, color treatments, and old comments from the client.
The rejected work mattered as much as the approved work. It showed what the client meant when they said “too busy,” “too generic,” “not enough product focus,” or “this looks like every other brand.”
The agent learned how the brand uses space, when the product should dominate the frame, which compositions feel premium, where text should sit, how much contrast is allowed, and what kind of generated background supports the message instead of stealing attention.
It also learned the practical details designers care about: platform dimensions, safe areas, Arabic and English text placement, campaign variants, export sizes, and what needs human approval before publishing.
From Prompt to Approval-Ready Artwork
The agent did not start from a blank canvas. It started by reading the campaign request and deciding what kind of visual post was needed.
For a product launch, it could create several image directions: a clean product-first post, a lifestyle-style generated scene, a comparison layout, or a story sequence. For an offer, it could prepare a stronger promotional layout with controlled hierarchy and clear pricing space.
It generated the image direction, arranged the post design, placed the product and text areas, adapted the same idea for feed and story sizes, and prepared variations for review.
The client could respond naturally: “make it quieter,” “more product focus,” “use the Eid campaign style,” “less template-like,” or “make this feel like the premium line.” The agent understood those notes because they were connected to real visual examples, not abstract adjectives.
What Changed
The design team did not stop designing. They stopped spending so much time fighting the first draft.
Instead of generating random AI images and trying to force them into a branded post, they worked with an agent that understood the client’s layouts, product treatment, visual boundaries, and approval habits.
First versions became closer to what the client wanted. Designers reviewed stronger directions, managers saw clearer options, and the client spent less time explaining the same visual preferences again and again.
The agent did not replace design taste. It carried the client’s visual taste into the workflow, so every new post started from a place that already looked familiar.