We Asked 100 Google Engineers How They Use Gemini: Here’s What We Learned

There is a lot of noise about AI in software engineering right now. Everyone has a hot take on whether AI is replacing developers or just making them lazy. But at Google, we don’t really rely on hot takes. We rely on data.

So, we did something simple. We went to our internal engineering directory, randomly selected 100 software engineers—from L3 (new grad) to Principal Engineer—and asked them one question:

“Show us the last prompt you sent to Gemini Code Assist.”

We didn’t want the marketing pitch. We wanted the raw, unpolished reality of how AI is actually being used inside the company that built it. What we found was surprising. It wasn’t just about generating boilerplate code. In fact, “writing code” wasn’t even the number one use case.

Here is the breakdown of how **Google engineers use Gemini** in 2025, and what you can learn from them to level up your own workflow.

Finding #1: The “Digital Archaeologist” (35% of Usage)

The most common use case wasn’t writing new code. It was understanding old code.

Every codebase has “that file.” You know the one. It was written five years ago by an engineer named “Dave” who left the company in 2022 to start a goat farm. It has zero comments, three nested loops, and a variable named `do_the_thing`. In the past, you would spend four hours staring at this file, trying to reverse-engineer Dave’s brain.

Today, Google engineers are using **Gemini Code Assist** as a translator.

Real Prompt Example:
“Explain this legacy Python script to me like I’m a Junior Java Developer. Specifically, what is the edge case handled on line 45?”

Why this matters: This is a massive productivity unlock. By using Gemini to perform “code archaeology,” our engineers are reducing their “time-to-understanding” by up to 50%. If you are interviewing for a role at Google, demonstrating that you can use AI to quickly grok an unfamiliar codebase is now a competitive advantage.

Finding #2: The “Agent Mode” Architect (25% of Usage)

In late 2025, we rolled out “Agent Mode” (preview) internally, and it has fundamentally changed the workflow for our Senior Engineers. Unlike standard autocomplete, which looks at your cursor, Agent Mode looks at your project.

Our survey showed a fascinating trend: Senior Engineers are using Gemini less for syntax and more for “multi-file orchestration.” They aren’t asking for a function; they are asking for a feature.

Real Prompt Example:
“I need to add a new ‘Dark Mode’ toggle to the settings page. Scan the codebase to find where user preferences are stored, update the schema, and generate the frontend toggle component in React. Don’t break the existing CSS.”

The Insight: The engineers who are thriving are the ones who treat Gemini like a smart intern. They delegate the “wiring” tasks—connecting the frontend to the backend—so they can focus on the system design. This is what we mean when we say AI makes you an “architect” rather than a “typist.”

Finding #3: “Vibe Coding” is Real (20% of Usage)

There’s a new term floating around our micro-kitchens: “Vibe Coding.” It sounds silly, but it refers to the practice of rapid prototyping using natural language where the “vibe” or the “intent” matters more than the specific syntax.

We saw this heavily among Product Managers and UX Engineers. They are using Gemini 3 Pro to spin up functional prototypes in minutes to test an idea before a single line of production code is written.

Real Prompt Example:
“Build a dashboard that looks like a 1980s sci-fi terminal. It should visualize server latency data. Use neon green text on a black background. Make it look ‘glitchy’.”

The Insight: This capability is closing the gap between “Idea” and “Demo.” If you are a non-technical candidate or a PM, learning to “vibe code” is the single highest-leverage skill you can acquire in 2025. It allows you to show, not just tell.

Finding #4: The Context-Switch Killer (15% of Usage)

One of the biggest killers of **developer productivity** is context switching. You’re coding, you hit a bug, you tab out to Stack Overflow, you fall into a rabbit hole, and 20 minutes are gone.

Our survey found that 15% of prompts were simply “lazy” questions—questions the engineer could have looked up documentation for, but didn’t want to leave their IDE to find.

Real Prompt Example:
“What is the gcloud command to list all active instances in us-central1 filtering by label ‘env=prod’? Just give me the command, no explanation.”

Why this matters: This isn’t laziness; it’s flow state preservation. By keeping the answer inside the IDE, engineers stay in “the zone” longer. The modern Google engineer treats Gemini as an infinite memory bank for syntax and CLI flags.

Finding #5: The Unit Test Skeptic (5% of Usage)

Finally, we saw a small but vocal group using Gemini strictly for what we call “The Vegetable Work”—the healthy stuff nobody likes eating. Writing unit tests.

Real Prompt Example:
“Here is a critical payment processing function. Generate 10 edge-case unit tests, including negative values, null inputs, and currency overflow. Be aggressive.”

The Insight: This is arguably the most responsible use of **Gemini for coding**. AI is fantastic at thinking of edge cases humans miss because humans are optimistic. AI is thorough. Using Gemini to pressure-test your own code is a hallmark of a Senior Engineer mindset.

What We Didn’t See

Interestingly, almost zero engineers in our sample used Gemini to “write the whole app” from scratch without supervision. The “replace the developer” narrative was nowhere to be found. The prompts were iterative, conversational, and highly specific.

How to Code Like a Googler in 2025

Based on this internal “survey,” here are three habits you should adopt if you want to align with **Google engineering culture**:

1. Stop Memorizing, Start Orchestrating: Don’t pride yourself on memorizing the entire standard library. Pride yourself on how fast you can stitch libraries together to solve a user problem.
2. Use AI for “Archaeology”: Don’t be afraid of legacy code anymore. Use Gemini to document code as you read it. It makes onboarding onto new teams incredibly fast.
3. Be a Prompt Engineer for Tests: If you aren’t using AI to generate your test suites, you are working too hard and covering too few edge cases.

The future of coding at Google isn’t less human. It’s more human. By offloading the syntax archeology and the test boilerplate to Gemini, our engineers are spending more time on the things machines can’t do: empathy, system design, and solving the really hard problems.

Want to try it yourself? Gemini Code Assist is available to developers everywhere. Start asking it the hard questions.

FAQ: Engineering with Gemini