# The AI-Assisted Coding Round Run, Rules Check to a Tagged Prompt Log

*Send one rules-check message, run a 60-minute AI-allowed round on a fixed clock, and log every prompt against the four criteria interviewers report grading.*

- Canonical URL: https://www.refolk.ai/candidates/guides/ai-assisted-coding-round-run
- Pillar: Interviewing
- Format: Playbook
- Published: 2026-11-09
- Last reviewed: 2026-11-09
- Reading time: 20 min
- Keywords: how to prepare for AI-assisted coding interview, AI-enabled coding round what to expect, can I use AI in my coding interview, how to use AI assistant during live coding interview, multi-file codebase coding interview practice, AI coding interview checkpoints time management

## Key takeaways

- Treat about 54 of the 60 minutes as yours, because candidates report the interviewer spends the first five to six minutes walking you through the platform, the files, and how to run tests.
- Candidates consistently report four grading criteria for AI-allowed coding rounds: problem solving, code quality, verification, and communication.
- AI permission applies to one round, not the whole loop: Meta reportedly pairs its AI-assisted round with a classic coding round that bans AI, and its own loop guide still asks candidates not to use AI tools.
- In Refolk's index only 497 US Software Engineers list GitHub Copilot as a skill, so the live round, not the profile, is where an interviewer sees whether you can direct an assistant.
- Run every change through prompt, review, run, confirm: explain a generated block before you paste it, then run the specific failing test before moving on.
- Running out of time in the third phase is survivable, since several reported candidates still got offers, but quickly accepted, unchecked AI output is the failure Canva names as Feature Marathon.

This is the run for a live coding round where the employer lets you, or expects you to, use an AI assistant inside a multi-file codebase. It works for any employer, and it covers loops where one coding round allows AI and another bans it. It gives you a one-message rules check, a 60-minute clock across the debug, extend, and optimize checkpoints, a prompt-and-verify loop the interviewer can watch, and a prompt log you write the same day.

## Which employers run AI-allowed coding rounds?

Canva and Karat NextGen describe their AI-allowed rounds on their own pages. Meta's and Google's formats are known only from candidate reports and secondary write-ups. Anthropic publishes the opposite default: no AI unless it says otherwise. If something does not appear on an employer's own page, treat it as unconfirmed until your recruiter confirms it in writing.

| Employer | Duration | Assistant | Outside AI | Source type |
|---|---|---|---|---|
| Meta | ~60 min | Built-in, model picker | Banned (reported) | Candidate reports |
| Canva | not established | Candidate's own | Allowed | Official blog |
| Karat NextGen | not established | Built-in | Prohibited | Official page |
| Google pilot | not established | Gemini required | unverified | Reporting |
| Anthropic | n/a | None unless told | Banned by default | Official page |

What each row actually rests on:

- **Meta.** Candidate reports describe a CoderPad environment with three panels: a file explorer, a code editor, and an AI assistant chat window. The round reportedly runs about 60 minutes, against 45 for Meta's standard coding rounds. The assistant can see the project files but answers only in the chat. Meta's own Full Loop Interview Guide, linked from its careers site, still asks candidates not to use unauthorized outside assistance, including AI tools like ChatGPT. The official document trails the reported practice.
- **Canva.** Canva says it now expects Backend, Machine Learning and Frontend engineering candidates to use AI tools like Copilot, Cursor, and Claude. For backend and frontend roles the round replaces its traditional Computer Science Fundamentals screening, because, in Canva's words, "AI assistants can trivially solve traditional coding interview questions."
- **Karat NextGen.** Karat's page describes a production-grade codebase, a VS Code-based IDE, and a built-in AI assistant. It says candidates are encouraged and often expected to use the integrated assistant, and that external AI tools are prohibited.
- **Google.** Only reporting so far, including the detail that the pilot requires Gemini.

This evidence base is thin. One of the most detailed public analyses of Meta's round draws on 35 candidate reports: 25 SWE, 9 MLE, and 1 backend. Formats change faster than write-ups do, so on the day you send your rules check, open the employer's own candidate page and note what it says. Your recruiter's written answer outranks that page and every third-party guide, this one included.

## What interviewers grade in an AI-assisted coding round

Candidates report four criteria: problem solving, code quality, verification, and communication. Meta has not published a rubric. Canva publishes three evaluation areas that cover the same ground, so you can prepare against all four criteria whichever employer you face.

Canva's three areas are Engineering Problem-Solving ("How do you approach ambiguous, realistic problems?"), Technical Depth & Ownership ("Do you understand and own the code in your solution?"), and AI Collaboration Effectiveness ("Can you critically evaluate AI outputs and iterate toward better solutions?"). The middle column of the table below maps Canva's areas onto the reported criteria. The mapping is my reading, not something either employer has published.

| Reported criterion | Closest Canva area | What proves it | When it lies |
|---|---|---|---|
| Problem solving | Engineering Problem-Solving | Plan spoken before the first prompt | Plan recited, then ignored |
| Code quality | Technical Depth & Ownership | Can explain every kept block | Tests pass, explanation stalls |
| Verification | AI Collaboration Effectiveness | Test run after each paste | Many pastes, few runs |
| Communication | Technical Depth & Ownership | Answers before asking the AI | Fluent narration of prompting, not of the problem |

The "when it lies" column is what an alert interviewer watches for. Any one of these criteria can be faked for a few minutes, and none of them can be faked across 50. The interviewer may also have more to go on than what you say. A third-party write-up reports that HackerRank lets interviewers watch the candidate's exchanges with the assistant live and captures the conversation in the interview report. Assume every prompt you type will be read.

**497 - US Software Engineers in Refolk's index who list GitHub Copilot as a skill**

That is 7.0x the 71 who list Cursor, and the Cursor figure may include database cursors, so treat it as unverified.

| Segment | Count | Top sample employer |
|---|---|---|
| US, GitHub Copilot | 497 | Microsoft (6 of 25) |
| Australia, GitHub Copilot | 38 | WiseTech Global (2 of 25) |
| US, Cursor | 71 | none above 1 |

In Refolk's index of professional profiles, the US has 13.1x as many Software Engineers listing Copilot as Australia, and Microsoft accounts for 24% of the US sample. My reading, which I have not verified, is that profile skill fields trail actual tool use. If so, an interviewer cannot see your AI fluency on your profile, and the live round becomes the main evidence. That is why the verification and communication habits below matter more than any prompting trick.

## The rules check: what to confirm before the round

Send one message at least a week before the loop. Ask, round by round, five things: where the assistant comes from, which model, which phases it covers, whether docs or search are allowed, and whether every other round defaults to AI off. Only the recruiter's answer settles these. Employer documents can lag behind what interviewers actually run.

Permission is layered, and each layer can be answered differently. A yes at one level tells you nothing about the level below it.

#### The layers an AI permission has to clear

1. **Loop** - Which rounds allow AI at all, and which default to off
2. **Round** - Built-in assistant or your own tool, and which interview it is
3. **Model** - Fixed model, a picker, or your choice
4. **Phase** - Whether AI is allowed for debugging and for direct questions
5. **Outside resources** - Documentation, web search, and any second tool

*A yes for the loop says nothing about a given round, and a yes for a round says nothing about outside tools or search.*

The best-sourced wording comes from a third-party template that asks the recruiter to confirm the AI policy for each round. It asks specifically whether any coding round includes a built-in assistant and whether any outside tool is allowed for the take-home. No employer publishes questions for this. Questions about model choice and docs or search access are not established employer guidance, so expect some recruiters to say they don't know. If yours does, ask the interviewer during orientation and write the answer down.

**Rules-check message to your recruiter**

```
Subject: AI policy for each round of my [role] interviews

Hi [recruiter name],

Thanks for setting up my interviews. Before I prepare, could you confirm the AI policy for each round?

1. Which rounds, if any, allow an AI assistant? For each one, is it built into the interview platform, or should I bring my own tool?
2. If it is built in, can I choose the model, and which models are available?
3. Is the assistant allowed in every phase of that round, including debugging and any questions the interviewer asks me directly?
4. Can I use documentation or web search during that round?
5. For every other round, including any take-home, should I assume no AI unless you tell me otherwise?

A one-line answer per round is perfect. Thank you,
[your name]
```

*Send at least one week out. Remove any line that a written employer page already answers, and keep the per-round framing.*

> **Rule:** Permission belongs to one round
>
> An AI-allowed round tells you nothing about the round after it. Unless the recruiter's written answer names a round as AI-allowed, treat it as AI off, take-homes included.

This default matches how employers write their own policies. Anthropic's guidance reads: "This is all you - no AI assistance unless we indicate otherwise," and for take-homes, "Complete these without Claude unless we indicate otherwise." Karat's permission applies to NextGen. It does not establish AI permission in every interview that carries the Karat name. Canva's AI round can fall in different places in the loop, and its own advice is to ask your recruiter which interview it is.

## How to split the 60 minutes

Plan on about 54 usable minutes. Orientation takes the first five to six, reading runs to minute 10, your spoken approach to 15, the build-and-verify loop to 50, and the defense fills the last ten. Test after every change rather than saving testing for one block at the end.

No employer has published a time split. Three public clocks exist, and they disagree in useful ways:

| Source | Read | Plan | Build | Test | Wrap |
|---|---|---|---|---|---|
| DG Learning | 0-5 | 5-15 | 15-40 | 40-50 | 50-60 |
| Hello Interview | interviewer orientation, 5-6 min | n/a | n/a | continuous | n/a |
| Canva frontend (third party) | requirements doc, 10-15 min | in requirements doc | n/a | ~45 | n/a |

There are two disagreements, and here is how I resolve them:

- **Testing as a block versus continuously.** The DG Learning clock puts testing at minutes 40 to 50. Hello Interview and TechScreen say to run tests immediately after each change. Go with continuous testing. An assistant you are not checking produces errors, and a test block at minute 40 finds them after you have built on top of them. The third-party Canva advice to have the AI add tests around minute 45 fits this: it adds coverage on top of the checks you ran after each paste.
- **Reading time.** One candidate was told to go straight to the tests with no reading time, while others were told to read first. Hello Interview's rule of thumb is "Five minutes reading saves fifteen minutes debugging." If the interviewer pushes you toward the tests, run the failing test so they see you comply, then say you are reading the files it touches before changing anything.

### The three checkpoints

Hello Interview reports three progressive phases built around one extended problem: fix a bug, implement the core feature, then optimize. In this guide those are the debug, extend, and optimize checkpoints. The danger is phase 1. One candidate reported spending half an hour on the first question, which is half the round gone before the main build begins. Before you start phase 1, decide how long you will give it and say that number aloud. If you hit it, tell the interviewer where you are stuck and ask whether to keep going or move on.

Not finishing phase 3 is survivable. Several reported candidates who ran out of time there still got offers. Canva says it would rather see thoughtful problem-solving and quality code than a rushed feature list. Spend your minutes on verified work in phases 1 and 2.

## The run, step by step

This is the full procedure from the rules check through the prompt log. Steps 1 to 3 happen in the week or more before the round, steps 4 to 7 happen inside the 60 minutes, and step 8 happens the same day.

#### The AI-assisted coding round, start to finish

1. **Send the rules check** - At least one week out, send the per-round message. Done when each round has an answer on assistant source, model, phases, docs or search access, and the AI-off default for the other rounds.
2. **Set up the environment** - Bring-your-own rounds: one to two hours getting your AI-integrated editor and a basic project in your language running. Built-in rounds: practise pasting from a chat-only panel. Done when editor, AI, and tests run end to end.
3. **Rehearse three timed runs** - On a multi-file codebase, run the tests early, then add one feature without breaking the existing tests. Do at least three full runs on the 60-minute clock.
4. **Orient and read the codebase** - Minutes 0 to 10. Follow the walkthrough, then read the data models, entry point, and failing test. Done when you can say what the data models are and what the failing test checks.
5. **Say your approach before any prompt** - Minutes 10 to 15. Explain your plan for the first checkpoint before typing anything to the assistant. Done when the interviewer agrees or redirects.
6. **Loop: scoped prompt, review, run, confirm** - Minutes 15 to 50. One narrow prompt tied to the plan, explain the output before pasting, run the targeted test, confirm aloud, move on. Repeat across debug, extend, and optimize.
7. **Defend the solution** - Minutes 50 to 60. Talk through the whole solution, what you rejected from the assistant and why, and what you would test next.
8. **Write the prompt log the same day** - About 20 minutes. List every prompt and correction with its minute, what you checked, and the outcome, each tagged to one of the four criteria.

### Setup that matches the round

Canva's prep list for bring-your-own rounds is short: "Have your preferred AI-integrated development environment ready; Set up a basic project structure in your preferred programming language." Pick the language the posting leans on. If you are unsure which one that is, [Refolk](/candidates) scores your fit against the posting and shows which required skills your history covers thinly, so your rehearsal time goes to the stack you will actually be tested on.

For built-in assistants such as Meta's and Karat's, setup is mostly rehearsal. You cannot install anything, so practise reading an answer in a chat panel and pasting it into a file yourself. That slower motion is exactly what gives you time to check, which the next section explains.

### Rehearsal that resembles the round

The drill comes from a practitioner who prepared for Meta's round: "Run the tests early. Add one feature without breaking the existing tests. Do this three times before your loop." Use a codebase with several files, an existing test suite, and at least one failing test. Single-function puzzle sites won't do, because a multi-file codebase is the whole point of this round. Record yourself, or ask a practice partner to interrupt you with questions, because the round grades how you talk through the work as well as the code.

## The prompt, review, run, confirm loop

Every AI contribution should pass through four visible moves: a scoped prompt tied to the plan you stated, a review where you explain the output before you paste it, a targeted test run, and a spoken confirmation before you move on. Hello Interview's summary of what interviewers reportedly want to see: "prompt, review, run, confirm, move on."

#### One pass of the verification loop

1. **Scope** - Name the file, function, and constraint, and ask for one small change
2. **Review** - Read the output aloud, check every symbol it uses exists in the codebase
3. **Paste** - Type or paste the change yourself, saying what you are keeping and dropping
4. **Run** - Run the specific failing or affected test, not the full suite
5. **Confirm** - State what passed, what you re-read, and what comes next

*Each paste happens only after the output has been read and explained, and the next prompt waits until a test has confirmed this one.*

### Scope the prompt

A scoped prompt names the file and function, says what the change must do, and gives any constraint from the plan you stated. Ask for one change at a time. A prompt like "implement phase 2" puts a whole design decision in the assistant's hands, and interviewers read that as Hands-Off. Expect the model to be limited. interviewing.io advises: "Expect a rather underpowered model, not anything state-of-the-art." Reported Meta pickers have mixed small models such as GPT-4o mini and Claude Haiku with larger ones such as GPT-5 and Gemini 2.5 Pro. Lists like this change, so confirm what is on offer during orientation and pick a model you have practised with.

### Review at the paste

A chat-only assistant gives you a natural checkpoint, because you type or paste every change yourself. Make the paste the moment you check, and do it where the interviewer can see. Say what the block does, name anything you are changing, and check that every method it calls exists. Given an underpowered model, unchecked output is likely to contain errors.

> The paste is your evidence of review. Skip the pause and you have only shown that the assistant can type.

### Run, then confirm

Run the test that targets the change first, then wider tests if time allows. After any edit made through a tool, re-read the file instead of trusting the tool's status message. Some candidates' starter prompts tell the assistant to re-read each file after a change, because the tool sometimes reports a timeout when the edit actually went through.

When the interviewer asks a question, answer before you prompt. Canva puts it plainly: "If your interviewer asks you a question, tell us what you think before getting help from AI." What counts as "verified" also varies by employer. Canva wants code you could submit as a pull request for peer review. Meta reports describe interviewers asking "Do you run code frequently? Do you check the AI's output before moving on?" Aim for the stricter standard of the two.

## How the AI-assisted round goes wrong

The expensive failures look like progress. Each one below comes with its false positive (what it looks like when it seems to be going well) and a check you can run during the round.

#### Progress against verification

Horizontal axis runs from Little verification to Every paste checked and tested. Vertical axis runs from Few checkpoints reached to Many checkpoints reached.

| Quadrant | What it means |
| --- | --- |
| Feature Marathon | Slow down, explain each block, and run a test after every paste |
| Directed and verified | Keep the rhythm and spend spare minutes on the defense |
| Hands-Off or Showcase | Stop prompting, restate the plan, and take back the design |
| Careful but slow | Survivable; cap phase 1 aloud and move on when you hit the cap |

*Fast progress without checks is the Feature Marathon, and slow but verified work is the quadrant that still gets offers.*

1. **Feature Marathon.** Canva describes it as accepting AI outputs quickly to maximize progress. *False positive:* many phases done, each shipped after a quick glance. *Check:* count how many pastes were followed by a test run. If the answer is not "all of them", slow down.
2. **Hands-Off or AI Showcase.** Hands-Off means treating the AI as a full solution generator. AI Showcase means prompting tricks crowd out the task. *False positive:* a fluent prompting session that never gets around to the actual problem. *Check:* does every prompt follow a plan you said out loud?
3. **"Tests pass" as proof.** The pattern reported: "The candidate pastes it into their editor, runs the tests, one passes, and they move on. When the interviewer asks what that function does, the candidate stumbles." *False positive:* a green test. *Check:* explain the code before pasting it, not after.
4. **Hallucinated APIs.** Reported failures include the assistant suggesting a method that does not exist in your codebase. *False positive:* code that reads plausibly. *Check:* search the codebase for every symbol the generated code uses.
5. **Full suite run too early.** One candidate said most sandboxes had about 1.5 CPU cores and 2 GB of memory, and that Meta's test suites are almost all stress tests. *False positive:* "I'm being thorough." *Check:* run the targeted failing test first and the full suite only at a checkpoint.
6. **Skipping the read because the interviewer said "run it".** In one candidate's words: "I jumped straight into the solution because he said 'let's run the test.' I should have taken a step back and read the whole code." *False positive:* looking responsive. *Check:* run the test, then read the files it touches before you edit.
7. **Assuming permission carries over between rounds.** In a split loop, an AI-allowed round sits next to one that bans AI. *False positive:* "they said AI was fine." *Check:* does the recruiter's written answer name this specific round?
8. **Trusting tool status messages.** *False positive:* a timeout or error after an edit that actually went through. *Check:* re-read the file before you retry the edit.

> **Watch out:** A green test is not ownership
>
> Passing tests show that the code runs, not that you understand it. If you cannot say what a function does before it goes into the file, you have not finished checking it, whatever the test says.

## The prompt log and the split loop after the round

The prompt log turns a blur of 60 minutes into evidence you can learn from. Within the same day, spend about 20 minutes writing down every prompt and correction, then tag each one to problem solving, code quality, verification, or communication. No published source documents this step. I recommend it because it shows you, in your own record, which criterion you neglected.

**Same-day prompt log**

```
Round: [employer, round name]   Date: [date]
Assistant and model: [built-in or own tool, model]
Rules confirmed by: [recruiter name, date of written answer]

Entry [n]
Minute: [approx.]
Checkpoint: [debug / extend / optimize]
Plan stated before prompt: [yes / no, and what it was]
Prompt: [verbatim or summary]
What I checked before pasting: [explanation given, symbols searched]
Correction made: [what I changed or rejected, and why]
Test run after paste: [which test, result]
Tag: [problem solving / code quality / verification / communication]

Totals
Pastes: [n]   Pastes followed by a test run: [n]
Questions answered before prompting: [n of n]
Checkpoint reached: [1 / 2 / 3]
```

*One block per prompt or correction. Write the prompt verbatim if you remember it, otherwise summarize it. Use exactly one tag per entry.*

Read the totals before anything else. If pastes outnumber test runs, your next rehearsal is about verification. If you prompted before answering a question even once, it is about communication.

### Keeping both modes sharp

A split loop needs two kinds of preparation in the same week. Candidates report that Meta pairs its AI round with one classic algorithm interview that bans AI. A secondary source adds that candidates at E7 and above get only one coding round, the AI-assisted one. So the mix depends on level. Label each round in your notes with its confirmed policy. Keep up unaided practice alongside your AI runs, so going without the assistant doesn't throw you in the banned round.

Line up your resume with what you demonstrated. If your resume lists Copilot or Cursor, the round is where that claim gets tested, and with only 497 US Software Engineers listing Copilot in Refolk's index, it is a claim that stands out. Refolk tailors your resume to each posting from your own history, so keep the AI-tool line only on applications where the round can back it up.

#### Before you call the round done

- [ ] The recruiter's written answer names which rounds allow AI, and every other round is logged as AI off.
- [ ] You know whether the assistant is built in or your own, and which model you will use.
- [ ] You have completed at least three timed 60-minute rehearsals on a multi-file codebase.
- [ ] In the round, you stated your approach before your first prompt.
- [ ] Every generated block was explained aloud before it was pasted.
- [ ] Every symbol the assistant used was found in the codebase before you relied on it.
- [ ] Every paste was followed by a targeted test run.
- [ ] You answered each interviewer question yourself before prompting.
- [ ] The prompt log was written the same day, with every entry tagged to one of the four criteria.

> **Tip:** Re-check the format before every loop
>
> These rounds are new and change quickly. Before each loop, reread the employer's own candidate page, ask the rules-check questions again, and log the date of the answer.

## Frequently asked questions

### Can I use AI in my coding interview?

Only if the employer has said so for that specific round. Anthropic, for example, states no AI assistance unless it indicates otherwise, and Meta's own loop guide still asks candidates not to use AI tools even though candidates report an AI-assisted round in the same loop. Ask your recruiter in writing, round by round, and treat any round they do not name as AI off.

### Should I bring my own AI tool to a Meta or Karat round?

No. Candidates report that Meta's round uses a built-in assistant inside CoderPad, and Karat NextGen's own page says external AI tools are prohibited and its integrated assistant is the one to use. Bringing your own tool is the pattern Canva describes, where candidates are expected to use tools like Copilot, Cursor, and Claude. Confirm which model applies to your round before you prepare.

### Do I need to finish all three phases to pass?

Reports say no. Hello Interview says several candidates who ran out of time in phase 3 still got offers, and a secondary source on Karat NextGen says candidates are not expected to finish every question. What sinks you is unverified code in the early phases. Canva says it would rather see thoughtful problem-solving and quality code than a rushed feature list.

### What do I do when the interviewer asks me a question mid-round?

Answer it yourself first, then use the assistant if you need to. Canva's guidance is direct: if your interviewer asks you a question, tell them what you think before getting help from AI. Going straight to the chat panel suggests you cannot reason without it. Give your view, say how you would check it, and only then prompt.

### Which model should I pick if the platform offers a choice?

Pick one you have practised with, and expect it to be weaker than the tools you use at work. Reported Meta pickers have mixed small and large models, and interviewing.io advises expecting a rather underpowered model, not anything state-of-the-art. Choose before the clock matters, say why you chose it, and check its output closely whatever you pick.

### How do I prepare when one round allows AI and another bans it?

Rehearse both modes in the same week and label each round in your notes. Candidates report that Meta pairs its AI-assisted round with one classic algorithm round without AI. Do your timed AI runs on a multi-file codebase and keep up unaided practice too, so going without the assistant does not throw you in the banned round.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/ai-assisted-coding-round-run*