Latest / The Tech Career Podcast with Fexingo: Engineering Jobs, Interviews, and FAANG Career Strategy / How FAANG Engineers Use GitHub Copilot in 2026
Transcript
- Lucas: So there's this moment that every FAANG engineer I know has experienced in the last year — you're staring at a blank editor, you type a comment like '// function to parse this JSON blob and return only the fields we need', and GitHub Copilot just... finishes it. And the weird thing is, it's usually right. Luna: And that's the part that freaks people out. Is it writing code you could have written, or is it writing code you wouldn't have thought of? Lucas: Right. And that's exactly the line we're going to walk today — how FAANG engineers are actually using Copilot in 2026, not just as a glorified autocomplete, but as a genuine productivity multiplier. But there's a catch. Luna: There's always a catch. Lucas: The catch is: you have to understand every line it generates. If you can't explain why a piece of Copilot code works, you should not ship it. That's basically the eleventh commandment at every FAANG right now. Luna: And honestly, if that insight was worth a coffee to you, there's a link in the show notes — buy me a coffee dot com slash fexingo. It's the smallest thing that keeps this show ad-free and independent. Lucas: Yeah, genuinely. Every cup helps us keep digging into how engineers actually work at these companies. So — back to that moment. I talked to a senior engineer at Meta last week who told me his team has an unspoken rule: Copilot is for scaffolding, not for core logic. Luna: Scaffolding meaning like, unit tests, boilerplate, data mocks? Lucas: Exactly. He said his team uses it to generate the first pass of unit tests — especially for edge cases they might forget. Like, they'll write a function signature, then prompt Copilot with 'write a test that checks what happens when the input list is empty, when it has one element, when it has a null value.' And Copilot will spit out a dozen test cases in seconds. Luna: But then a human has to review each one. Because Copilot doesn't know your internal API's error handling conventions. Lucas: Exactly. And that's where the review process is actually getting more rigorous, not less. Because now you're not just reviewing the code a human wrote — you're reviewing code a human chose to accept. So the bar for 'why did you let this through' is higher. Luna: I've heard some teams at Google actually require engineers to use Copilot for certain repetitive tasks. Like generating data access layers or serialization code. Lucas: Yeah, I've heard that too. There's a team in Google Cloud that mandates Copilot for any code that maps between protobuf and JSON. It's too error-prone to do by hand, and Copilot's training data includes enough protobuf examples that it gets it right 95 percent of the time. Luna: But what about the other five percent? Lucas: That's the code review. And that's the skill FAANG is now hiring for — not just 'can you write code,' but 'can you evaluate ai generated code critically.' I saw a job posting from Amazon last month that literally listed 'experience with ai assisted development tools' as a preferred qualification. Luna: Amazon — which is famously cautious about AI in production. They banned Copilot internally for a while, right? Lucas: They did. In 2023, they blocked it over data security concerns — they didn't want proprietary code being sent to GitHub's servers. But by 2025, they rolled out their own internal version — basically a Copilot clone that runs entirely inside AWS, trained on Amazon's own codebase. Luna: So it's less about 'should we use AI' and more about 'whose AI do we trust.' Lucas: Exactly. And that's where the conversation has shifted. Every FAANG now has some flavor of internal Copilot. Even Apple has one — they call it 'Swift Assistant', it's only for Swift and Objective-C code. Luna: So for an engineer interviewing at FAANG right now, is it worth learning how to prompt Copilot well? Lucas: I'd say it's becoming table stakes. Not because you'll use it in the interview — you won't, they lock down the browser — but because the interviewers themselves use it every day, and they're looking for candidates who think in that mindset. One of the Meta engineers I talked to said he asks a question now: 'If you had Copilot open on this problem, how would you prompt it to get started?' Luna: That's a fascinating lens. It's not about the answer — it's about how you'd decompose the problem for a language model. Lucas: Right. And there's a specific prompt framework engineers are using. It's called 'context, constraints, format.' You give Copilot the context — here's the function, here's the data structure. Then the constraints — this has to be thread-safe, no external dependencies, must handle nulls. Then the format — return a JSON object with these fields. Luna: And does that actually produce better results than just saying 'write a parser'? Lucas: Dramatically. I tested it. I gave Copilot a vague prompt — 'write a function to validate an email address' — and it gave me a basic regex. Then I used the framework: 'Context: we have a user registration form with fields email and confirmEmail. Constraints: must handle Unicode, must reject disposable email domains, must be case-insensitive. Format: return a boolean and an error message string.' The second output was production-ready. Luna: So the skill is really about specificity. Breaking a vague requirement into concrete, testable pieces. Lucas: Exactly. And that's the same skill that makes a good software engineer in general. So Copilot isn't replacing engineering judgment — it's amplifying it. Luna: But it's also creating new failure modes. I've heard stories of engineers who blindly accepted Copilot code that introduced subtle bugs — like off-by-one errors in loops, or incorrect variable shadowing. Lucas: Oh, absolutely. There's a famous internal post at Google about an engineer who shipped a Copilot-generated function that handled date formatting. It worked perfectly for nine months, then broke on February 29th of a leap year. Because Copilot's training data didn't have enough leap-year examples. Luna: That's terrifying. And exactly why code review can't be automated away. Lucas: Right. So the companies that are getting the most value out of Copilot are the ones that have built strong review cultures. They treat Copilot like a junior engineer — it produces a lot of code fast, but everything it writes needs a senior eye. Luna: And that changes the dynamics of code review. Instead of reviewing a small change from a human, you're reviewing a large change that was generated, and you have to be more skeptical. Lucas: Exactly. One Amazon engineer told me his team now requires every Copilot-generated block to be tagged with a comment — 'ai generated, reviewed by ' — so there's accountability. Luna: That's smart. And it also creates a paper trail for postmortems when something goes wrong. Lucas: Yeah. And I think that's actually the most interesting thing about this whole shift — it's making engineering culture more explicit. The norms that used to be unwritten are now being codified because AI forces you to be clear about what you expect. Luna: So what's the takeaway for someone listening who wants to work at FAANG? Should they learn to use Copilot, or avoid it? Lucas: Learn it. But learn it critically. Use it in your side projects. Pay attention to when it gives you wrong answers — that will teach you more about your own blind spots than any coding challenge. Luna: And maybe mention in interviews that you use it. That shows you're current. Lucas: Yeah. But be ready to talk about a specific example where you caught a mistake it made. That's the story that demonstrates real engineering judgment. Luna: I like that. It's not about being the person who uses AI — it's about being the person who uses AI wisely. Lucas: Exactly. And that's the bar FAANG is setting now. So if you're prepping for interviews, spend some time with Copilot. But spend more time reviewing what it writes. Luna: Alright, that's a solid note to end on. Thanks Lucas. Lucas: Thanks Luna. See you next time.