Don't Use Claude for Everything: The People Who Actually Make AI Workflows Work Are Building Systems

Amir Mušić tweeted that he had used Claude to complete a full brand project.

Competitive research, reference gathering, content strategy, design, content planning, ad banners—he went through the whole thing. In the end, even the brand itself was built with Claude's help.

In the context of AI tools, this sounds like a demonstration of how capable Claude is. But the line he paired it with points in the opposite direction: Stop trying to use Claude for everything.

Don't treat Claude as an all-purpose player.

Taken together, these two lines mean: Claude can perform well on specific tasks, but if you push the entire workflow onto it, you're heading in the wrong direction. He followed up with something more direct: Think wider — build a setup, not just agents.

In other words, build a system, don't just pile up agents.

This idea is worth unpacking. Most people still use AI by “opening a chat window and dumping the whole problem into it.” That's like hiring a jack-of-all-trades employee but giving them no team, no tools, and no historical records. They might do the task at hand well, but next time the task changes, they still won't know who you are or what you want.

One reply in the quoted tweet put it more clearly: the memory matters more than the 22 pages. No matter how polished the 22-page output is, if the system has no memory, next time you're starting from zero.

A setup, when you break it down, really comes down to three things.

First, role separation. Competitive research stays research, design stays design, content stays content. Each agent has clear boundaries; you don't ask one agent to pretend it understands everything.

Second, a memory mechanism. Archive what gets produced each time, then pull it back up and keep using it next time. This matters far more than which model you choose.

Third, workflow orchestration. From research to strategy to design to output, there is an order; it's not all thrown into one pot.

Under the tweet, someone said “bookmark this guys, huge time saver,” and someone else commented directly, “I immediately wanted to buy eggs with that branding.” The latter actually shows one thing: the result itself is concrete enough that people have a real reaction to it, rather than thinking, “stuff made by AI is just okay.”

The food branding design shown in the video is the result of this workflow. Four dishes, four lines of copy—“Dip in.”, “Roll with it.”, “Five more minutes.”, and “Stay a while.”—paired with corresponding images, clean and simple. Looking good is the surface; the reusable process underneath is the key.

For people with a technical background, this approach isn't complicated; the hard part is actually building the role and memory relationships between agents. For non-technical readers, at least remember one thing: don't expect a single chat window to solve every problem. Break tasks apart, fill in the context, and save each output. The results will be much better than repeatedly trying things in one window.

Back to the source material itself. Amir Mušić's tweet didn't go into the specific architecture of the setup. He mentioned, “this setup works too well to skip it,” and said he is using it himself and with clients. He didn't fully explain what it looks like. But the direction is already clear.

Rather than looking for an AI that can do everything, build a structure that can keep doing the work.