AI-assisted coding interviews: what companies say

Three companies have said, in public, that you may use AI in the room. I read what each one says. Here is what they say, and what I do with it.

What the three companies say

Canva. In June 2025, Canva wrote that it expects backend, machine learning and frontend candidates to use AI tools. A pilot competency called AI-Assisted Coding replaced its computer science fundamentals screen for backend and frontend roles. Canva lists what it evaluates. The list includes when to use AI and how to break down vague requirements. It also includes catching and fixing problems in AI-generated code. Canva tells candidates to practice with these tools first.

DoorDash. In March 2026, DoorDash described a 60-minute AI-assisted working session. It says the session replaces traditional coding rounds. You work on your own machine, in your own editor. You share your screen and narrate. The five things DoorDash names are getting oriented, using AI and verifying it, debugging, managing scope, and communicating tradeoffs. I read that post line by line in a separate piece.

Meta. Hello Interview reports that Meta began rolling out an AI-enabled coding round in October 2025. It runs 60 minutes in CoderPad with a built-in assistant. It replaces one of the two coding rounds at the onsite. You work one extended problem in three phases: fix bugs, build the core, then optimize. The four competencies are problem solving, code quality, verification and communication. Finishing every phase is not required.

What they share

Read the three again. Verification shows up every time. Canva wants you to catch bad AI output. DoorDash wants you to verify it. Meta lists verification as its own competency.

Talking shows up too. DoorDash asks you to narrate. Meta lists communication.

So here is my reading. The skill is not typing prompts. The skill is reading code you did not write, deciding whether it is right, and saying why.

How I work the room

This is my method, built from those sources and from interviewing.io's advice on Meta's round.

1. I read the task before I open the assistant. I say what I think it asks, out loud. Coco's three questions are a good start.
2. I hand the assistant a small job with a clear edge. A helper function. A list of test cases. I keep the design for myself.
3. I read the answer before I run it. I predict what it will do. Then I run it.
4. I hunt for the case the assistant skipped. Empty input. Repeats. Order.
5. I say what I accepted, what I changed and what I threw away.

Here is the fourth step in action. Say the assistant hands you this for "remove repeats and keep the first one of each, in order."

def dedupe(items):
    return list(set(items))

It looks right at a glance. A set has no promised order, so it breaks the "in order" rule. I say that out loud, then fix it.

def dedupe(items):
    seen = set()
    result = []
    for item in items:
        if item not in seen:
            seen.add(item)
            result.append(item)
    return result

The fix took a few lines. The sentence I said before it is the one that counts.

What weak use looks like

interviewing.io describes the weak version plainly. The candidate treats the assistant as an end-to-end solver. They paste output they can't explain. They write a vague prompt, then re-prompt again and again. They let the assistant change large chunks of code in one go.

The strong version is the reverse. Work in small steps that you can review. Check each step before you build on it. When the assistant's assumptions are unclear, ask the interviewer. A clarifying question costs little and saves you from a wrong turn.

Practice the part that stays yours

You cannot rehearse every company's setup. DoorDash lets you bring your own editor. Meta uses CoderPad. Setups differ. The habits don't.

You can rehearse reading code, predicting what it does, and explaining it. Pick a function from an old project. Say what it does before you run it. Then run it and find out. StrongYes does not recreate any company's round. It trains the narrating.

Coco is an AI study partner, not a person. She listens while you explain your code and asks one question at a time. She never hands you the fix.

The editor and the code runner are free. Open the editor and talk through a problem you already solved. Contains Duplicate is a small one to start with.