I'm an Amazon tech lead who uses AI to write code daily. There's one situation I hesitate to use it in.

• An Amazon tech lead says it's hard to resist the allure of vibe coding.
• Anni Chen, who vibe codes daily, says it's definitely a productivity boost.
• But she cautions against using vibe coding at scale or for production, and says technical knowledge matters.

This as-told-to essay is based on a conversation with Anni Chen, who has worked at Amazon for about three-and-a-half years. It has been edited for length and clarity. Business Insider has verified her employment history.

I'm a tech lead at Amazon responsible for deploying large-scale generative AI and LLM-driven systems. I focus on what we call memory, which powers personalization in generative AI experiences across Amazon.

I vibe code every day. It's definitely a productivity boost.

For debugging or small tasks, I sometimes treat it like a lottery. Maybe it will produce something amazing. Sometimes, it does.

Vibe coding helps me brainstorm what the solution could look like, even if I don't adopt the final solution it proposes. Vibe coding also speeds up the time spent rewriting code when you realize a requirement wasn't covered.

Vibe coding is faster, even with double-checking

When I vibe code, it's always iterative. I give it the basic information it needs, it produces a version, then I check it — similar to a code review with coworkers. I might say, "You missed this part" or "You missed that part."

The AI sometimes fixes issues but introduces something new. You have to keep an eye on it.

For complex tasks, you need more double-checking. But even with the extra checking, it's still faster.

I was working with a partner team and ran into complex locking issues. Without an LLM, I might have taken a day to research possible solutions, especially since it was relatively new to me.

Within 15 minutes, I brainstormed with the LLM about possible solutions. I pointed out weaknesses in its suggestions and asked it to improve them. In 15 minutes, I had a proposal to send to the team.

Technical knowledge helps — you know what's a good solution and what's not. You know what tastes good, but you don't know what dishes are available. The LLM brings up all the dishes, and you choose.

Vibe coding for scale might not work, and technical knowledge still matters

Still, I'm hesitant to use vibe coding directly in production.

LLMs are very good at solving problems, but sometimes they make implicit assumptions you don't realize they're making. If you don't tell it explicitly, for example, that something needs to work for multi-threading, it might just produce the minimum version that works, but when it's large-scale or productionized, it could crash.

Non-technical builders could tell an LLM to build something that handles millions of users. But if you have zero technical knowledge, it's hard to anticipate constraints upfront. If you don't tell the model the implicit assumptions, it won't respect those constraints. Later, you'll run into problems.

Non-technical people might use the LLM to fix issues reactively. But technical people can anticipate constraints proactively and prevent problems in the first place.

Technical people also understand vibe-coded content better, and they're in a better position to understand what LLMs are good at and not good at. For example, knowing how they're trained and why they're weaker at certain tasks like math. That understanding helps you master them as tools.

When you scale to one million or 100 million customers, systems need to be coded differently to handle that scale.

Vibe coding is hard to resist

Initially, leadership pushed vibe coding. Our team is a GenAI team, so we were naturally more receptive. In non-GenAI teams, engineers initially reacted like, "No, I won't let AI do my job. I don't trust AI-generated code."

After people tried it, attitudes shifted. People realized it's pretty good sometimes. Now it's more widely adopted.

It's very hard to resist vibe coding nowadays. If you're an employee, leadership sees the productivity boost and will encourage you to use it.

When your peers are using it and coding faster, it's hard to resist. If you can't keep up with the speed, it becomes difficult to collaborate.

Even if you resist, you still consume AI passively. AI comments are embedded in code reviews. So even if you don't vibe code directly, you're still interacting with AI outputs.

Do you have a story to share about vibe coding? Contact this reporter at cmlee@businessinsider.com.