By FDE Interview Prep, a SourceStrongAI company

In this post13 sections
  1. The short answer
  2. What they can see when the assistant is allowed
  3. Before you prompt: scope the problem out loud
  4. Prompting in front of an interviewer
  5. Verify before you accept
  6. When the model is wrong
  7. Working in an existing codebase with an assistant
  8. Mistakes that make the assistant look like the candidate
  9. Which interviews allow an assistant
  10. Practice this before the day
  11. Questions people ask
  12. Keep reading
  13. More from the blog

The interviewer just told you the assistant is allowed, and your first thought is the obvious one: if the model can write the code, what are they grading? The answer is you: what you ask for, what you check, and what you refuse to accept. This post gives you the words to say at each step, a verify-before-accept habit, and a plan for the moment the model is wrong. It covers one round; the FDE interview guide covers the whole loop.

The short answer

In an AI-assisted coding interview, the code is cheap, so your judgment is what shows. Scope the problem out loud before you prompt. Prompt in small, checkable steps. Read and test every suggestion before you keep it, and say why you kept it or threw it away. When the model is wrong, catch it on a small case, name the mistake, and fix it.

Employers that describe these rounds point the same way. Leo Mehr, Ramp’s Director of Engineering, said in June 2026 that Ramp wants to see how candidates prompt Claude, think about the model they use, and scope the problem. Source 1What is a Forward Deployed Engineer? (FDE Explained) feat. Leo Mehr of Ramp (YouTube auto-generated English captions)PublisherdearCC (Clara Shih), YouTubeSource typerecorded talk or interview Sierra says its AI-native onsite makes it easier to gauge a candidate’s agency (do they pivot when stuck?) and judgment (how do they scope what to build in the time?). Source 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog

What they can see when the assistant is allowed

The interviewer cannot see inside your head. In a live round they see four things, and in a take-away build they see the same four afterward, through what you show and explain:

  1. What you do before the first prompt. Did you ask questions and state a plan, or paste the problem in?
  2. What you ask for. A small, specific request shows you know what you want. “Solve this” shows you don’t.
  3. What you do with the output. Did you read it, run it, and test it, or accept it and move on?
  4. What you say. Your narration is the only view they have of your reasoning.

A public GitHub repo from an organization named Axium Industries describes a live interview in which the interviewer plays the customer and the candidate builds a small with any coding agent. Source 3Second-stage interview — Forward Deployed Engineer (AI)PublisherAxium Industries (GitHub organization)Source typecompany hiring page It lists what is scored: customer instincts, agent architecture, judgment and verification, and communication. Source 3Second-stage interview — Forward Deployed Engineer (AI)PublisherAxium Industries (GitHub organization)Source typecompany hiring page That is one organization’s repo, not a standard, but notice that none of those dimensions is “wrote the code”.

The rest of this post is our method for making those four things visible. It is how we coach, not an employer’s rubric.

Before you prompt: scope the problem out loud

Your strongest signal comes before the model does anything. Start the way you would without an assistant: restate the problem, ask about inputs and edge cases, and say what “done” looks like.

Say it like this:

  • “Before I prompt anything, let me restate it: given a list of booking windows, return them merged where they overlap.”
  • “Two questions. Is the input sorted? And if one booking ends exactly when the next one starts, is that an overlap or two back-to-back bookings?”
  • “Here’s my plan. I’ll have the assistant draft the merge, I’ll write three test cases myself, and I’ll check the edge cases before I trust it.”

Suppose the interviewer answers: “Unsorted. Back-to-back bookings stay separate.” You now know two things the model does not. Write three tests before you prompt: an out-of-order pair, two touching windows, an empty list.

That last line matters most. It tells the interviewer that the assistant works for you, and that you already know how you will judge its output. Sierra says it tells candidates before its onsite that it is fine to cut scope while building and to skip boilerplate such as CRUD and auth to focus on what is unique. Source 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog Scoping out loud is how you show you know which part is unique.

Prompting in front of an interviewer

Meta says candidates for select roles use an authorized AI assistant inside CoderPad during technical interviews. Source 4Meta Jobs Hiring ProcessPublisherMetaSource typecompany hiring page In a shared editor like that, assume the interviewer can see your prompts, and write them as if they are reading over your shoulder.

Prompts that show judgment

  • Ask for one piece at a time: a function, a test, a parser. Not the whole solution.
  • Put your decisions in the prompt: “Input may be unsorted. End times are exclusive.”
  • Name the constraints: the language version, no new dependencies, match the existing style.
  • Ask for tests or edge cases separately, so you can compare them with your own.
  • Keep the prompt short enough to read aloud.

Narrate each prompt in a sentence: “I’m asking for just the merge function, and I’ve told it the input may be unsorted, because that’s the case I expect it to miss.”

If the interviewer steps out

Not every format has someone watching. Sierra says that in its Build phase the interviewer steps out while the candidate builds with the tools of their choice. Source 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog Narrating to an empty room does nothing; what counts there is how you explain your decisions afterward.

Our method: keep a running decisions log in the README as you build. One line each for what you cut and why, what you rejected from the model and why, and the bug you caught and how. When the interviewer comes back, walk the review through that log instead of trying to remember.

Verify before you accept

This is the habit that separates you from the tool. Every suggestion gets read, run and tested before you keep it. Here is the loop we teach:

  1. Read it aloud, briefly. “It loops once, extends the last window when they overlap, else starts a new one.”
  2. State what it assumes. “It assumes the input is sorted by start time.”
  3. Test on a small case you chose. Pick the case you expect to break, not the one in the prompt.
  4. Decide, and say why. “Keeping it with one change” or “Rejecting it, because...”.

A worked example. You asked for the merge, and the assistant returns this:

def merge(windows):
    out = []
    for s, e in windows:
        if out and s <= out[-1][1]:
            last = out[-1]
            last[1] = max(last[1], e)
        else:
            out.append([s, e])
    return out

It reads well, and the sorted example from the prompt passes. Now run the case you expected it to miss:

merge([(5, 8), (1, 6)])
# returns [[5, 8]]; expected [[1, 8]]

It never sorts, so an earlier booking that arrives second gets swallowed. It also treats windows that only touch as overlapping, because of <=. The interviewer told you back-to-back bookings stay separate, so that is wrong too: merge([(1, 3), (3, 5)]) returns [[1, 5]].

The fix is two small edits: iterate over sorted(windows), and compare with <.

def merge(windows):
    out = []
    for s, e in sorted(windows):
        if out and s < out[-1][1]:
            last = out[-1]
            last[1] = max(last[1], e)
        else:
            out.append([s, e])
    return out

The three tests you wrote before reading its code now pass:

assert merge([(5, 8), (1, 6)]) == [[1, 8]]
touch = [(1, 3), (3, 5)]
assert merge(touch) == [[1, 3], [3, 5]]
assert merge([]) == []

Say the close out loud: “Sorted input and exclusive ends. Both edge cases pass, so I’m keeping it.” The interviewer just watched you find two bugs the model wrote with confidence. That is the signal.

Don't trust the assistant's own tests

If you ask the model for tests after it writes the code, its tests tend to share its assumptions: the unsorted input it forgot in the code, it will likely forget in the tests too. Write at least the edge-case tests yourself, before you read its code.

When the model is wrong

It will be, and that is good news. A wrong suggestion is your best chance to show judgment. What matters is how you handle it.

Say what you are doing. “This looks off. Let me trace it on a small input before I change anything.”

Name the mistake precisely. Not “it’s buggy” but “it loses a booking when the input arrives out of order.”

Choose fix or re-prompt, and say why. If the fix is a line or two, make it yourself: “Faster to fix than to explain.” If the model misunderstood the task, re-prompt with the failing case: “Here’s an input where your version returns the wrong answer. Rewrite it to handle unsorted input.”

Know when to stop prompting. If two re-prompts have not fixed it, write it yourself. “The model keeps missing the same case, so I’ll write this part myself.” Looping on prompts while the clock runs shows the opposite of judgment. Sierra says it looks at whether candidates pivot when they get stuck. Source 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog

Other ways models go wrong in interviews:

  • APIs that are gone or never existed. df.append(row) looks right, but current pandas no longer has DataFrame.append(); the fix is pd.concat. Run it before you build on it, and say “Let me check this call exists in the version we have.”
  • Silently changed behavior. You ask it to tidy a function, and the “cleanup” renames user_id to userId in the returned dict, which breaks every caller. Diff it against what you had: “It changed the output keys; I’m reverting that part.”
  • Confident complexity claims. It says “this is O(n log n)”, but it calls sorted() inside the loop. Check it yourself: “The sort is inside the loop, so it runs once per item. I’ll sort once, up front.”

Working in an existing codebase with an assistant

Some AI-enabled rounds are described as starting from existing code. The Blind poster asking about the Cohere FDE onsite wrote that the recruiter described an existing Python codebase and AI. Source 5Cohere FDE onsite, AI enabled coding round?PublisherBlindSource typecandidate report on Blind Sierra says it is piloting a debugging interview in which candidates improve a colleague’s draft PR in a medium-sized codebase using coding agents, and that the level of AI allowed there is still to be decided. Source 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog

Reading code you did not write is the core skill here; the free question on refactoring a long function so you can test it drills it. The trap in this format is prompting before reading. The model does not know the codebase’s conventions, so it invents new ones. Do this instead:

  1. Read first. Run the tests, find where your change goes, and name the pattern to copy: “Errors here raise FilterError, and the handler maps it to an HTTP 400.”
  2. Write the acceptance test yourself. Include the case you expect the model to miss.
  3. Prompt with the pattern. “Add a date filter to this handler, raise FilterError on bad input the way parse_limit does.”
  4. Own the diff. Be ready to explain any line it wrote, and revert anything you did not ask for.

Our free lesson on what FDE coding rounds test covers the other coding formats you may meet. For writing the test before the fix, the Pro question write a failing test that reproduces a reported bug is good practice.

One person’s view of where this is heading: Tom McLaughlin, a founding forward deployed engineer at the OpenAI Deployment Company, said on MBN Solutions’ “A Class Act” podcast in September 2026 that for FDE hiring he would assess how well candidates orchestrate agents, build feedback loops and build their own tools. Source 6The Role of The Forward Deployed Engineer with Tom McLaughlin, FDE at The OpenAI Deployment CompanyPublisherMBN Solutions (YouTube)Source typerecorded talk or interview The episode says guest views are their own, not the employer’s.

Mistakes that make the assistant look like the candidate

  • Silent prompting. In a live round, you type, it answers, you accept, and nobody spoke. Fix: narrate one sentence per prompt and one per decision. If nobody is watching, the decisions log does this job.
  • Accepting the first draft. Fix: nothing is kept until it has run on a case you chose.
  • Outsourcing the questions. Asking the model what the edge cases are before asking the interviewer. Fix: the interviewer is the customer; ask them first.
  • Using a tool that isn’t allowed. Meta says no outside AI tools are authorized, only those in its interview environment. Source 4Meta Jobs Hiring ProcessPublisherMetaSource typecompany hiring page Fix: ask the recruiter, and use only what they name.
  • Reading AI answers aloud. GitLab asks candidates not to read from AI-produced answers in interviews. Source 7Using AI in the interview processPublisherGitLabSource typecompany hiring page Fix: use the assistant for code, and do your own talking.
  • Refusing to use it. If the employer expects you to use the assistant, as Meta says of select roles, avoiding it hides the very thing they want to see. Source 4Meta Jobs Hiring ProcessPublisherMetaSource typecompany hiring page Fix: use it, and show your judgment in how.

The talk about AI may not end with the code. Practice the Pro question how do you use AI coding tools in your daily work, and where do you not trust them?, and answer with the habit you just showed: where you trust the tool, where you check it, and a time it was wrong. If the assistant question comes up for a take-home instead, read can you use ChatGPT or Claude on a take-home?

Which interviews allow an assistant

Rules differ by company and by stage, so check each one before the day. Three examples:

Reports from candidates are thinner: one completed FDE report, one invited candidate who did not name the role, and one second-hand question asked before an onsite. One OpenAI Forward Deployed Engineer candidate, in an Aced report whose listing dates the interview to May 2026, described a separate AI-enabled coding screen as “basically an easy LeetCode problem”. Source 10OpenAI Forward Deployed Engineer Interview ExperiencePublisherAced (formerly Exponent)Source typecandidate’s personal write-up A Decagon candidate invited to an onsite wrote on Blind in May 2026 that it had an “AI Coding” round they thought focused on building an app using AI; the poster did not name the role. Source 11Decagon - AI Coding Interview (Blind)PublisherBlindSource typecandidate report on Blind A Blind poster asking for a friend with an upcoming Cohere FDE onsite wrote, in September 2026, that one round is AI-enabled coding on CodeSignal of about 45 minutes, with “an existing python codebase and ai”; the onsite had not happened yet. Source 5Cohere FDE onsite, AI enabled coding round?PublisherBlindSource typecandidate report on Blind

The full list, by company and stage: which companies allow AI tools in interviews, and the Pro lesson on AI assistance rules by company and stage turns it into a list to check against your targets. If your recruiter has not said which tools you may use, ask.

Practice this before the day

The narration feels awkward the first time, so do it before the interview, not in it.

Your practice session

  • Pick a coding question and set a timer.
  • Before any prompt, say your restatement, two questions and your plan aloud.
  • Write two edge-case tests yourself before reading any generated code.
  • After each suggestion, say what it assumes and whether you keep it.
  • Find one bug the model wrote, and say how you found it.
  • Record yourself and listen back for silent stretches.

Run the checklist above on the Pro question extend a feature in a codebase you have never seen, using an AI assistant: it has the follow-ups an interviewer asks, such as why passing assistant-written tests is not enough, and a model answer to check yourself against. Pro starts with a 7-day free trial. If you would rather rehearse scoping out loud first, the free practice case needs only a sign-in.

GlossaryForward deployed engineerA software engineer who builds and ships production systems inside a customer’s problem and environment, accountable to that customer’s outcome.More on Forward deployed engineerGlossaryAgentA system in which a model chooses steps and tool calls to complete a task, within limits the design sets.More on Agent

Questions people ask

What do interviewers look for in an AI-assisted coding interview?

Two employers that describe these rounds point to how you work with the tool. Ramp’s Director of Engineering said he wants to see how candidates prompt Claude and scope the problem, and Sierra says its AI-native onsite makes it easier to gauge a candidate’s agency and judgment.Source 1What is a Forward Deployed Engineer? (FDE Explained) feat. Leo Mehr of Ramp (YouTube auto-generated English captions)PublisherdearCC (Clara Shih), YouTubeSource typerecorded talk or interviewSource 2The AI-native interview (Vijay Iyengar, Arya Asemanfar, Angie Wang)PublisherSierraSource typecompany blog

Should I use the AI assistant if the interview offers one?

Yes, when the employer expects it. Meta’s hiring page says candidates for select roles are expected to use the authorized assistant in CoderPad, and that no outside AI tools are authorized.Source 4Meta Jobs Hiring ProcessPublisherMetaSource typecompany hiring page

What should I do when the AI assistant gives me wrong code?

Say you are checking it, run or trace it on a small case, name the mistake, then fix it yourself or re-prompt with the failing case. Catching the error out loud shows the verification habit the interviewer cannot otherwise see.

Can I use my own AI tools in the interview?

Only where the employer says so. Meta says only the tools in its interview environment may be used, and OpenAI’s interview guide says expectations for AI tools vary by interview and that candidates who are unsure should ask their recruiter. Ask before the day which tools you may use.Source 4Meta Jobs Hiring ProcessPublisherMetaSource typecompany hiring pageSource 8Interview guidePublisherOpenAISource typecompany hiring pageSource 9OpenAI interview guide (Wayback Machine capture)PublisherOpenAI (archived by Internet Archive)Source typearchived company page

Keep reading

Guides

Lessons

Questions

More in Interview rounds