Clients Are Using AI to Tell Us What They Want to Build, Pay For, And Ship Without Ever Talking to Us

Picture this…

You've got a CEP plugin holding your entire production pipeline together. Then Adobe announces it is sunsetting CEP and your dev team just quoted you six months to migrate to UXP. Later that day, someone in leadership asks the obvious question: can AI just rebuild it?

That question followed me to NAB 2026.

Adobe pulled together about 50 third-party plugin developers for a private session. Product, Engineering, and Partner teams in the same room. The official agenda was UXP for Premiere and After Effects. What's coming, what's missing, where it's headed.

Then the Q&A session took a turn.

Adobe wanted to know how developers were building plugins with AI. They went around the room and asked. I heard almost unanimously positive responses.

Comments like: active AI assistance. Full agentic vibe-coding. Developers described AI doing the majority of their software development.

Then came the question nobody expected: was anyone NOT using AI for production code?

I raised my hand. The room laughed. The kind of laughter that says thank goodness someone said the quiet part out loud.

I explained that Hyper Brew doesn't use AI for production code. We write the final code by hand. It's faster and more reliable for the kind of complex Adobe plugin development we do. Where correctness isn't optional.

The room went quiet for a few awkward moments.

Then two other developers raised their hands and said the same thing. After the session a few more people came up to me ( including a few Adobe employees ) who expressed a similar frustration at the misleading hype around vibe-coding, grateful that someone had pushed back.

Best use for AI in Plugin Development

The most useful thing AI has done for our process isn't on our end. It's on the client's.

A client vibe-codes a rough version of what they want. It half-works. It's janky. But it shows us exactly what they're trying to build. We take the prototype, throw out the code, and build the real version from scratch.

What used to take multiple calls and long spec documents now takes a ten-minute demo.

A client came to us with a vibe-coded prototype as a jumping off point. The code was unusable. The idea was crystal clear. That prototype gave us a clear vision of what the client needed.

That's AI doing what it's good at right now: lowering the cost of communicating a vision. Not replacing the people who build it.

The importance of quality production code cannot be overstated.

The tools our clients depend on need to work during a live broadcast, a client presentation, the busiest week of the year. That's not where you want to find a ‘vibe’ nobody caught.

If an AI-generated image goes wrong, someone ends up with extra fingers. If AI-generated code goes wrong, the consequences are far worse. Files get deleted. Websites go offline. Client data gets leaked. It's not a risk we're willing to take just to save a few hours.

AI has real, logical use cases we rely on every day.

We use it for research, code examples, image recognition, audio to text, and more. But writing production code is not one of them. Not today. The tools aren't reliable enough, and the consequences of getting it wrong are too serious.

We stand by the code we write. That's why we write it ourselves.

The Misconception We Keep Running Into

We've had some companies assume we're vibe-coding everything. That the plugins we build are just cleaned-up AI output with our name on it.

We're not.

Every line of production code is written by a human who knows Adobe's APIs, understands how creative workflows function under pressure, and can be held accountable when something breaks.

That's why we added a line to our marketing: Tools made by humans for humans.

Prototype with AI. Just don't ship the prototype.

Where This Goes Next

We'll stay on top of new technologies no matter their hype. AI tools will change. But right now, they're not a reliable choice for production code.

Where they do make sense, we use them. Research. Prototypes. Non-LLM applications like image recognition, audio to text, classification, and rotoscoping. Real use cases with real results.

For production code? We write every line ourselves. That's not stubbornness. That's accountability.

For now: we build what works. We stand behind it. And we write every line ourselves.

Need a plugin built right the first time?

Talk to us at hyperbrew.co/booking