By FDE Interview Prep, a SourceStrongAI company

In this post11 sections
  1. Why you have to build the pressure yourself
  2. The kit: a timer, an unseen prompt, a recording and a scorecard
  3. Decomposition mock: where to get prompts and how to score yourself
  4. Client round mock: a script for the friend who plays your customer
  5. Coding mock: practice in the tool you will use
  6. Behavioral mock: record, transcribe, cut
  7. Using AI to prepare, and where employers draw the line
  8. Reviewing the recording: what to fix first
  9. Questions people ask
  10. Keep reading
  11. More from the blog

You have a (FDE) loop coming up, and nobody to practice with. The loop won’t give you a practice round, so the pressure you will feel on the day has to be built at home. This post is the practice half of the FDE interview guide, which covers the rounds themselves.

The short answer: an FDE mock interview is a timed, recorded rehearsal of one round, run on a prompt you have not seen, scored afterwards against a checklist you wrote before you started. For the client round you add a person, or an AI, playing the customer from a script. If you want one tonight with no friend free, the free practice case runs a 10-minute decomposition and client mock against an AI customer, with a free sign-in, and scores it.

Below: the kit, a setup for each round, a script for the friend who plays your customer, and what to fix first. The checklists and the script are our method, not any employer’s rubric.

Why you have to build the pressure yourself

Reading a model answer tests recognition. The round tests saying the right thing, in order, out loud, while someone watches and the clock runs. Only the first improves by reading.

Palantir’s recruiting blog, in an August 2022 post for intern and new-grad candidates, told them to think out loud in the technical interview and to share questions and observations so interviewers could see how they approach the problem. Source 1From Pipeline to Prospect: Insights and Advice from Palantir Recruitment (Palantir Blog; archived copy)PublisherPalantir Technologies (Palantir Blog on Medium)Source typearchived company page Palantir also tells onsite candidates to talk through how they plan to approach analytical and technical problems, and to ask questions if necessary. Source 2Palantir Careers | Getting HiredPublisherPalantirSource typecompany hiring page You cannot practice either of those silently.

A mock rebuilds three pressures:

  • The clock. You have to commit before you feel ready.
  • The unknown. You do not know the prompt, or what the customer is hiding.
  • The audience. Someone, or a recording, hears every pause and every “um, so, basically”.

Drop any one of them and the mock turns back into studying.

The kit: a timer, an unseen prompt, a recording and a scorecard

PieceHow to set it up
TimerA phone timer you can see. Never pause it.
Unseen promptChosen by someone else, or drawn at random from a list you wrote days ago.
RecordingVoice memo for talk rounds; screen and voice for coding and whiteboard rounds.
ScorecardWritten before you start, filled in only while you listen back.

Four rules make it a mock and not a rehearsal:

  1. No restarts. If you open badly, recover out loud, as you would on the day.
  2. Talk the whole time. When you are thinking, say what about.
  3. Stop when the timer stops, and give your close anyway.
  4. Score from the recording. Memory is kind to you; the recording is not.

How long should the timer run? Use the length your recruiter gives you for that round. If you do not have one, practice at more than one length, because a plan that fits a long round can fall apart in a short one. Our full building permits case lets you pick a 30-minute, 45-minute or 60-minute clock.

Decomposition mock: where to get prompts and how to score yourself

The gives you a vague customer problem to break down into something buildable.

Where to get prompts you have not seen

  • The question bank. Take the city bus reliability question and read only the prompt line: a transit agency says its buses are unreliable and riders are leaving. Close the tab before you see the framework.
  • A public dataset. Palantir’s own preparation advice for open-ended questions includes building real things, such as finding a public dataset and thinking of something interesting to do with it. Source 3Palantir Careers | Navigating Open-Ended QuestionsPublisherPalantir TechnologiesSource typecompany hiring page Turn that into prompts: open a city’s open data portal, find something messy (service requests, inspections, permits) and write one sentence as the customer would say it. Put it in a list and leave it for a few days, so you forget the details.
  • An AI model. Snowflake’s candidate guidance suggests using AI to generate sample interview questions. Source 4AI Cheat Sheet: How (and When) to Use AI in Your Snowflake InterviewPublisherSnowflakeSource typecompany hiring page Ask for problems, not answers:
Write six one-sentence problems a
customer might give a forward deployed
engineer. Mix government, healthcare,
logistics and retail. Each should sound
vague, mention a goal or a complaint,
and hint at data the customer already
has. Do not suggest solutions.

How to run it

Open a blank page or a whiteboard app and split it into lanes: clarifications, stakeholders and metrics, inputs, workstreams, first version, risks. Start the timer, read the prompt once, and say your first question out loud before you write anything.

Here is the difference in the first minute, on a prompt you won’t be using as a mock: a clinic group says “our no-shows are killing us”.

Weak opening: “I’d build a model that predicts which patients will miss their appointment and text them a reminder.”

Strong opening: “Before I design anything: when you say no-shows, do you mean patients who never arrive, or ones who cancel the same morning? And is it every clinic, or a few? Each points at different data.”

The weak opening is not wrong about the tech. It is wrong about the order: it answers a question nobody has asked yet.

Score it

Palantir’s page on open-ended questions advises articulating the alternatives and trade-offs, being pragmatic enough to reach a concrete approach, and delivering a functioning idea first, then expanding it. Source 3Palantir Careers | Navigating Open-Ended QuestionsPublisherPalantir TechnologiesSource typecompany hiring page Our checklist turns that into things you can hear on a recording:

Decomposition mock scorecard (ours)

  • My first sentence was a question, not a solution.
  • I pinned what the vague word means in something you can count.
  • I named who decides, who uses it, who pays and who could block it.
  • For each data source I asked who owns it, how fresh it is and how I get access.
  • I ordered the work by risk and said which order I rejected, and why.
  • I described a first version a real user could use soon, and what it fakes.
  • I named failure modes, including one about people rather than tech.
  • I closed with the problem, the first version, the biggest risk and what happens on Monday.

Our lesson The open-ended round and the method on one page walks through each step on a worked case, and it is free.

Client round mock: a script for the friend who plays your customer

The client round is the hardest one to practice alone, because it needs a second person. A friend with no engineering background is fine, since the customer they play isn’t an engineer either. What they need is a script, so they hold facts back instead of helping you.

The card to hand your friend

Here is a card to start from. Better still, let your friend change the hidden facts, so you walk in not knowing them.

WHO YOU ARE: head of operations at
a regional grocery chain.
WHAT YOU WANT: "an AI forecast so we
stop throwing away fresh food."
YOUR BOSS: promised the board a demo.

HIDDEN FACTS (give one only when a
question points straight at it):
- Most waste is in bakery and produce.
- Store managers override the order
  system by hand; nobody records why.
- The sales data arrives a day late.
- IT has refused two vendors before.

DURING THE CALL:
- Answer in one or two sentences.
- If they pitch a tool before asking
  about the waste, say: "You haven't
  asked where the waste happens."
- Once, get annoyed: "We tried this
  before."
- Once, ask: "Can you promise me a
  date?"
- Never summarize for them.

Tell your friend the most important rule twice: a vague question gets a vague answer. “Tell me about your data” earns “we have lots of data”. “Where does the order quantity come from, and who can change it?” earns the manual-override fact.

What to say first

A strong opening sets an agenda and asks before proposing:

“Thanks for making the time. I’d like to spend the first part understanding where the waste happens and what you’ve tried, then suggest a small first step, and end with who does what next. Does that work? To start: where does most of the waste show up?”

When your friend pushes back (“We tried this before”), acknowledge the specific point before you move: “That’s useful. What happened last time, and what would have to be different for you to trust it?”

When they ask for a promise, commit only to what you control, with a date: “I can’t promise the forecast by then, because I haven’t seen the data. I can promise that by Friday you’ll have a one-page plan and an answer on whether the sales feed is usable.”

Score it, and ask your friend three questions

Use the moves from our free lesson on what customer rounds look like, and the core moves: an agenda, diagnosis before solution, a playback of what you heard, acknowledgment before disagreement, options with a recommendation, dated commitments you control, and a close with owners and dates.

Then ask your friend, before they have time to be polite:

  • Did you feel understood? Where did you stop feeling that way?
  • What exactly did I promise you?
  • Would you take a second meeting with me?

If their answer to the second is not what you meant to promise, your words were loose.

For a harder variant, answer the CTO who must hear the integration will land late, with your friend playing the CTO. Our post on the customer role-play interview has more exchanges to rehearse.

Coding mock: practice in the tool you will use

A coding mock fails when the practice setup is kinder than the real one, so find out the tool first. Anthropic’s careers page says that for technical roles it uses live coding tools such as Colab and CodeSignal, and that you can look things up but should be comfortable with basic syntax and standard libraries. Source 5CareersPublisherAnthropicSource typecompany hiring page Palantir’s phone interview page says it may ask you to use HackerRank or Google Docs. Source 6The Phone Interview - Palantir CareersPublisherPalantir TechnologiesSource typecompany hiring page

Then match it:

  • A shared doc. Write the code in a plain document, with no autocomplete and no way to run it. Trace one example by hand, out loud.
  • A notebook or an online editor. Practice there, not in your usual editor with its plugins.
  • AI completion. If the round will not have it, turn it off for the mock.

Take the per-user token bucket rate limiter question as your prompt. Run the mock before you read on, and say your approach before you type.

The lesson here is how you test. Pass the time in instead of sleeping, so every result is exact and you can say it before you run it:

b = Bucket(cap=2, rate=1, now=0)
calls = [b.allow(0) for _ in range(3)]
assert calls == [True, True, False]
assert not b.allow(0.5)
assert b.allow(1.0)

Narrating that test is the point: “Two tokens to start, so the third call fails. Half a second later there’s half a token, still not enough. At one second there’s a whole one.” Then do the question’s second half, making it work across several processes, without looking anything up.

On the recording, check that you asked about edge cases, stated the approach before coding and named its cost. Our free lesson on what FDE coding rounds test covers the other formats you may meet.

After your mock: a minimal version to compare

Tokens refill at a set rate up to a cap, and a request goes through only if a whole token is there to take:

class Bucket:
    def __init__(self, cap, rate, now):
        self.cap = cap
        self.rate = rate
        self.tokens = cap
        self.last = now

    def allow(self, now):
        elapsed = now - self.last
        self.last = now
        gain = elapsed * self.rate
        t = self.tokens + gain
        self.tokens = min(self.cap, t)
        if self.tokens >= 1:
            self.tokens -= 1
            return True
        return False

The per-user part is a dictionary keyed by user id, with a bucket created on each user’s first request:

buckets = {}

def allow_user(user, now):
    if user not in buckets:
        new = Bucket(2, 1, now)
        buckets[user] = new
    return buckets[user].allow(now)

Behavioral mock: record, transcribe, cut

Behavioral answers get better by editing, and you cannot edit what you have not recorded.

  1. Record it cold. Take the one-minute background question, which ends with why this role, and answer with no notes.
  2. Transcribe it with your phone or any transcription tool.
  3. Find where the answer starts. Highlight the first sentence only you could say, and delete everything before it.
  4. Cut the “we”. Each “we” in a story about your work hides what you did. Make it “I” or cut it.
  5. Check the ending. The last sentence should answer the question asked, not trail off.
  6. Record it again from the edited transcript, then once more without it.

Here is what the cutting steps do to a real opening:

Before: “So, um, I’ve been in software for a while, mostly backend, and we built a lot of data pipelines at my last company...”

After: “At my last job, an insurer, I rebuilt the claims pipeline so adjusters saw new claims the same morning instead of the next day, and I did it by sitting with them first.”

The cut removed the warm-up and the “we”, and kept a sentence only you could say.

The transcript is where AI helps most. Paste it in with a request like this:

Here is a transcript of my spoken
answer to "tell me about yourself"
for a forward deployed engineer role.
Point out filler, any sentence where
"we" hides what I did, and any claim
with no concrete result. Do not
rewrite it for me.

Keep the words yours: an answer someone else wrote falls apart at the first follow-up.

Using AI to prepare, and where employers draw the line

Employers in both kinds of FDE hiring encourage AI for preparation, and say where it stops.

The practical rule for your mocks: rehearse each round under the rule of the real round. Use AI freely to generate prompts, play a customer and critique your transcripts. If the real round bans AI, run the mock without it, or you are rehearsing a crutch.

AI makes a decent stand-in customer when no friend is free, and it can keep the secrets from you too:

Play a customer in a job interview
role-play. You run a regional grocery
chain and want "an AI forecast" to cut
food waste. Invent four hidden facts
about where the waste happens and why
(data, people, IT, a past failure).
Do not tell me them. Reveal one only
when my question points straight at
it. Answer in one or two sentences.
Push back once. Never suggest
solutions or summarize for me.

When the call ends, ask it to list the four facts, and score how many you found. That number goes on your scorecard.

It will still be more helpful than a real customer, so treat any fact it volunteers as a mock that went easy on you. The free practice case at /try is built for this: its AI customer, a city whose mayor has already promised an AI review tool, reveals facts only when your question targets them, and the score quotes your own words back to you.

Reviewing the recording: what to fix first

Listen back once with the scorecard in front of you. Then pick one thing to fix, in this order, because each problem hides the next:

  1. You proposed before you asked. Everything after a premature solution was built on a guess. Fix: make your first sentence a question, every time.
  2. No first version. You explored for the whole round and never committed. Fix: say “here’s the first version” out loud before the halfway mark.
  3. Nothing concrete. No data sources by name, no fields, no rough numbers. Fix: name the tables and one field that matters.
  4. Silences. Mark each stretch where you stopped talking. Fix: say what you are weighing, even when you have not decided.
  5. No close. The timer ran out mid-thought. Fix: rehearse the close on its own until it comes out whole.

Our Pro lesson on narration, the board and the clock goes deeper on the last two. Write the fix at the top of your next scorecard, and run the next mock on a new prompt, not the same one: a prompt you have seen twice tests memory. Our post on common FDE interview mistakes lists the fix for each round’s traps, and if the loop is days away, how to spend the days before your onsite sets out which practice to run on which day.

Your first mock can start tonight. Run the free practice case cold, with no notes. When the score arrives, check it against the eight lines of the decomposition scorecard above, pick the first item from the fix-first list that you missed, and write it at the top of tomorrow’s card.

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 engineerGlossaryDecomposition roundAn open-ended interview in which you work out loud from a vague problem with several possible solutions to a concrete approach and a first working version.More on Decomposition round

Questions people ask

Can I use ChatGPT or Claude to practice for an FDE interview?

Several employers encourage AI for preparation. Anthropic encourages candidates to use Claude to research the company and practice answers, and Snowflake suggests using AI to generate sample interview questions. Microsoft adds that during assessments and interviews candidates should demonstrate their own skills unless explicitly permitted.Source 4AI Cheat Sheet: How (and When) to Use AI in Your Snowflake InterviewPublisherSnowflakeSource typecompany hiring pageSource 7Guidance on Candidates' AI UsagePublisherAnthropicSource typecompany websiteSource 8Hiring tipsPublisherMicrosoftSource typecompany hiring page

How do I practice a decomposition interview alone?

Use a timer, a prompt you have not seen, a recording and a checklist you score afterwards, and talk out loud the whole time. Palantir’s own preparation advice for open-ended questions includes building real things, such as finding a public dataset and thinking of something interesting to do with it.Source 3Palantir Careers | Navigating Open-Ended QuestionsPublisherPalantir TechnologiesSource typecompany hiring page

Is there a free FDE mock interview I can try?

Yes. The free practice case on this site runs the building permits case against an AI customer after a free sign-in, and scores your run. It works as a mock for the decomposition and client rounds.

Keep reading

Guides

Lessons

Questions

More in Preparation