Forward deployed engineer take-homes: what real prompts ask for and how to scope yours

By FDE Interview Prep, a SourceStrongAI company

In this post12 sections

1. The brief says build something useful; it means more than code
2. What company-published prompts ask for
3. What candidate-reproduced briefs ask for
4. The pattern: a small build, and the work around it
5. Scoping your time box: the cut list and the stub list
6. What to write instead of building
7. Preparing for the walkthrough and the live extension
8. Mistakes that cost the most
9. Before you submit
10. Questions people ask
11. Keep reading
12. More from the blog

The recruiter’s email says the next step is a take-home, and the brief fits on one screen: build something useful for a fictional client in a few hours. You are already picking a framework, which means you are about to spend the whole time box on code. Beside the code, the forward deployed engineer take-home briefs we read keep asking for three things: writing addressed to the client, a record of how you used AI, and a walkthrough where you defend what you built. Not every brief asks for all three, so read what yours asks you to submit before you read the problem. For every other round in the loop, start with the forward deployed engineer interview guide.

The brief says build something useful; it means more than code

Read the “what to submit” section before the problem. That is where the other deliverables hide.

Tangible’s published brief is the clearest example: a 2-hour timebox with three parts: a feed validation and re-run-safe ingestion service for a fictional client’s positions CSV, a follow-up email of 200 to 400 words to the client’s integration lead, and a short AI usage note. ^[1]Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany website

Ethyca’s take-home for the Forward Deployed Privacy Engineer role says it is not a programming interview. ^[2]Source 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website

Spend the whole box on code and you hand in one part of three.

What company-published prompts ask for

These briefs sit in public GitHub repositories under the companies’ names; AzurePartners is an account by that name, unconfirmed as a company.

What the table leaves out:

• Coframe gives you 2 hours, and the video covers a demo, your solution with every AI tool you used, edge cases and productionization. ^[3]Source 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany website
• Ethyca says its challenge should take between 3 and 4 hours. ^[2]Source 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website After you submit, it walks through your solution with you on an interview call. ^[2]Source 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website
• Namastex writes its brief in Portuguese, gives about 3 calendar days, and points the agent at a deliberately unreliable quote API. ^[4]Source 4Desafio Técnico — FDE / AI Engineer (Namastex)PublisherNamastex (GitHub)Source typecompany website
• AzurePartners asks you to aim for about three hours, no more than four, within a 7-day deadline. ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website Part A, worth 70 points, is graded against a larger hidden set of contract tests, and shortlisted candidates get a 45-minute session to extend the build live. ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website
• LevelUp Labs files its brief under fde-take-home-challenge, but the prompt never says FDE. ^[7]Source 7LevelUp Labs Take-Home Challenges: EngineeringPublisherLevelUp Labs (levelup-labs-ai on GitHub)Source typecompany website
• ELITH asks you to use two reference customers to decide what belongs in the shared product foundation, and to explain it in an FDE Decision Memo. ^[6]Source 6ELITH LOCK: FDE Challenge 2026PublisherELITH (elith-co-jp on GitHub)Source typecompany website

What candidate-reproduced briefs ask for

These come from candidates’ own GitHub repositories. Read each as one person’s report, not the company’s current process.

• Kong. One candidate reported, in a repo created in September 2026, a “Kong PS / Forward Deployed Engineer” exercise: a live kickoff call of 30 to 45 minutes where the interviewer plays a customer stakeholder, a bounded async build targeting about 4 to 6 hours, and a live readout and demo of 45 to 60 minutes. ^[8]Source 8Kong PS / Forward Deployed Engineer — Take-Home ExercisePublisheralexscottan (GitHub)Source typecandidate’s take-home repository
• Tavily. One candidate reported, in a repo created in August 2026, a brief that expects roughly 4 to 6 hours and offers two options: improve a starter agent CLI, or write a technical explainer. ^[9]Source 9Tavily - FDE - Take Home AssignmentPublisherazamkhan99 (GitHub)Source typecandidate’s take-home repository
• Adobe. One candidate reported, in a repo created in February 2026, a brief that says “Plan to spend 3-4 hours MAX”, then 30 to 35 minutes of presenting and a live demo and edit with engineers. ^{[10, 11]}Source 10FDE Take-Home Exercise: Creative Automation for Social CampaignsPublishernew-marty (GitHub)Source typecandidate’s take-home repositorySource 11creative-automation-pipeline (repository description)Publisherjtdman (GitHub)Source typecandidate’s take-home repository
• Concourse. One candidate reported, in a repo created in September 2026, a brief that supplies a deliberately messy account export and asks which of “50-some” accounts need attention. ^{[12, 13]}Source 12Forward Deployed Engineer: Take-Home (PDF)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repositorySource 13vibe-coding (repository description)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repository

The pattern: a small build, and the work around it

This is our reading of the briefs above, not a count.

A working build, smaller than you want

Tangible’s brief asks for “a complete, smaller solution” over “an ambitious fragment”, and the Tavily brief one candidate reported says “We’d rather see a small thing done well than a large thing done loosely.” ^{[1, 9]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 9Tavily - FDE - Take Home AssignmentPublisherazamkhan99 (GitHub)Source typecandidate’s take-home repository The AzurePartners brief would “rather read a focused submission with honest gaps than a padded one”, and the Concourse brief one candidate reported says “A defensible answer on rough plumbing beats perfect plumbing and no answer.” ^{[5, 12, 13]}Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany websiteSource 12Forward Deployed Engineer: Take-Home (PDF)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repositorySource 13vibe-coding (repository description)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repository

One instruction: finish something small and say what you left out.

Writing addressed to the client

Tangible wants an email to the client’s integration lead, and Ethyca an email reply to the customer whose API error you diagnosed. ^{[1, 2]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website The Concourse brief one candidate reported asks for a short readout. ^{[12, 13]}Source 12Forward Deployed Engineer: Take-Home (PDF)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repositorySource 13vibe-coding (repository description)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repository

Write it to the person the brief names, not to the reviewer. Writing for customers covers the follow-up email line by line.

A record of how you used AI

Tangible asks for an AI usage note, Coframe for every AI tool you used, Namastex for your exported AI conversations, ELITH for an AI Usage Memo, and LevelUp Labs for the prompts or sessions you used. ^{[1, 3, 4, 6, 7]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany websiteSource 4Desafio Técnico — FDE / AI Engineer (Namastex)PublisherNamastex (GitHub)Source typecompany websiteSource 6ELITH LOCK: FDE Challenge 2026PublisherELITH (elith-co-jp on GitHub)Source typecompany websiteSource 7LevelUp Labs Take-Home Challenges: EngineeringPublisherLevelUp Labs (levelup-labs-ai on GitHub)Source typecompany website The Tavily brief one candidate reported asks for a record such as coding-agent chat history or session logs. ^[9]Source 9Tavily - FDE - Take Home AssignmentPublisherazamkhan99 (GitHub)Source typecandidate’s take-home repository

Tangible’s brief gives the shape: which tools you used, what you used them for, and one place where you overrode or corrected what the AI gave you. ^[1]Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany website An illustrative note in that shape, for a fictional brief:

• Tools: Claude Code for the CSV parser and its tests.
• Used for: first drafts of the validators. I wrote the dedup rule myself.
• Overrode: it keyed duplicates on row number. I changed the key to account, fund and as-of date, because the client re-sends files in a different order.

The override line shows you were driving. Using AI on a take-home covers what to disclose when the brief is silent, and how you use AI tools is the question you will get about it in the walkthrough.

A walkthrough you defend

Coframe wants it recorded. ^[3]Source 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany website Ethyca walks through your solution with you on a call, and the AzurePartners brief has shortlisted candidates extend the build live. ^{[2, 5]}Source 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany websiteSource 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website Alexey Grigorev’s AI Engineering Field Guide, which covers take-homes for AI and ML engineering roles generally, describes a “defence round” in which candidates present and defend their solution in a 45–90 minute walkthrough interview. ^[14]Source 14AI Engineering Field Guide — Home assignmentsPublisherAlexey Grigorev (GitHub)Source typeindependent analysis

Scoping your time box: the cut list and the stub list

The cut list is not our invention. The AzurePartners brief’s first DECISIONS.md question is “What did you deliberately not do, and why?”, and the Concourse brief one candidate reported asks for a cut list: “What you decided not to do, and why.” ^{[5, 12, 13]}Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany websiteSource 12Forward Deployed Engineer: Take-Home (PDF)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repositorySource 13vibe-coding (repository description)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repository The time split and the stub list are ours.

Split the box before you open an editor. Tangible’s brief does the split for you: about 60 to 75 minutes for the service, 20 to 30 for the email and roughly 5 for the AI note, inside 2 hours. ^[1]Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany website When your brief gives no split, budget the writing and the walkthrough first, and give the build what is left. Here is a plan for a fictional brief. It follows Tangible’s proportions, not Tangible’s numbers:

Brief: 4 hours. Build, email, README.
0:00 read brief, write README heads
0:15 skeleton: input to output
1:15 riskiest part (messy rows)
2:30 stop building, commit
2:45 client email
3:15 README: cut, stub, AI, time
3:45 clean clone, say walkthrough

Build the thinnest path end to end. Input in, output out, ugly in the middle. A walking skeleton that runs beats three polished modules that don’t connect.

Order the rest by risk. Build the part most likely to break first: the messy input, the flaky API, the ambiguous field. Namastex’s evaluation list asks what the agent does when the quote API fails, and calls that “the point that most separates” (our translation from Portuguese). ^[4]Source 4Desafio Técnico — FDE / AI Engineer (Namastex)PublisherNamastex (GitHub)Source typecompany website Sequencing by risk shows the method on a bigger project.

Ask, or decide and write it down. Tangible tells candidates to email questions “as you would Dana”, its fictional integration lead, and Ethyca invites you to reach out and ask. ^{[1, 2]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany website The AzurePartners brief says you may instead make a call and write it down, but “silently guessing and not saying so is not” fine. ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website One candidate reported a Concourse brief that says to ask if you’re blocked on logistics, while “Judgment calls about the data are yours.” ^{[12, 13]}Source 12Forward Deployed Engineer: Take-Home (PDF)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repositorySource 13vibe-coding (repository description)Publishercgillespie529-gif (GitHub)Source typecandidate’s take-home repository

A question worth sending is short, states your default, and can be answered in one line:

Hi Priya, two rows in the visit export share a visit ID but list different clinics. I’m treating the later row as the correction and flagging both in the report. Tell me if your export works the other way.

If no answer comes, keep your default and put it on the Assumptions line of your README.

Keep two lists as you go. They are different things, and a reviewer needs to tell them apart.

• The cut list is what is not in this submission, with one line on why. “No authentication: the brief names one internal user, and the reviewer runs this locally.”
• The stub list is what exists as a shape but fakes its behavior. The code path is real; the call at the end is not.

A stub should announce itself. Here is one for a fictional brief where a clinic’s visit export feeds a scheduling system you have no credentials for:

def push_visit(visit: dict) -> dict:
    """Stub: no scheduler credentials.
    See README, 'Stub list'."""
    vid = visit["id"]
    print(f"[stub] push {vid}")
    return {"id": vid,
            "status": "stubbed"}

It prints [stub] push V-1042 and returns a record marked stubbed, so no one mistakes it for a real integration. The stub list entry says what the real version needs: retries, and an idempotency key so a re-run never books the same visit twice.

What to write instead of building

Everything you cut becomes a sentence, and your README carries most of them:

The client note is shorter and plainer. A skeleton for the fictional clinic brief:

Hi Priya, the visit export now loads into a staging table on every run, and re-running it does not create duplicates. Two things need your call before we go live. First, some rows have no clinic code; I have held them in a review list rather than guess. Can your team tell me which clinic they belong to? Second, I have not connected to the scheduler yet, because I don’t have credentials. Once I do, it is a small change, and I will send a test run for you to check before any real visit moves.

It says what works, what is blocked and who decides.

For the sentence you will need most, borrow this shape: “I did not build X. With the time I had, I chose Y, because Z. Here is what X would take.”

Preparing for the walkthrough and the live extension

Coframe’s list of what the video should cover makes a good default outline for any walkthrough: a demo, the solution and the AI tools you used, edge cases, and what it would take to run in production. ^[3]Source 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany website Recording the walkthrough video goes into pacing and what to show on screen.

Prepare answers to the three questions a walkthrough keeps coming back to:

• “Why this slice first?” Point to the risk you retired. “The input was the least certain part, so I made sure we could load every row before building anything on top.”
• “What breaks first in production?” Name one thing. “Retries against the scheduler, because the stub hides them. I’d add a dead-letter table and alert on it.” If “dead-letter” is new to you, the dead-letter queue entry explains it in a paragraph.
• “What would you change with another day?” Answer from your README. Practice it with this question from our bank, which also asks what you left out on purpose.

If your brief includes a live extension, as the AzurePartners one does, leave a seam for it. ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website Keep the path from input to output in small, well-named functions so a new field touches one place. Before the call, rehearse one extension end to end: add a new field, carry it through, and show it in the output.

Mistakes that cost the most

1. Building the whole brief and writing nothing. Fix: draft the README headings and the client note’s first line before you write code, then fill them in at the end.
2. Going over the time box and hiding it. The AzurePartners brief says “Please don’t spend more than four.” ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website Fix: stop at the limit, and state your time in the README.
3. Treating the visible tests as the goal. The AzurePartners brief says “The visible tests are a floor, not a ceiling.” ^[5]Source 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website Fix: test the edge cases the brief implies, not only the ones it shows.
4. Guessing the AI rule. Namastex expects AI coding tools, while Anthropic asks candidates to complete take-homes without Claude unless it says otherwise. ^{[4, 15]}Source 4Desafio Técnico — FDE / AI Engineer (Namastex)PublisherNamastex (GitHub)Source typecompany websiteSource 15Guidance on Candidates' AI UsagePublisherAnthropicSource typecompany website Fix: find the sentence in your brief, and ask the recruiter if there isn’t one. The AI rules by company and stage lesson lists the published rules.
5. Silent stubs and silent guesses. A fake that looks real reads as a bug or a bluff, and so does an assumption nobody wrote down. Fix: label every stub in the code, and list every stub and assumption in the README.

If your take-home is for an AI lab, the same routine applies; the OpenAI FDE take-home shows it against what OpenAI publishes and what candidates report.

Before you submit

The Kong brief one candidate reported opens with a kickoff call where the interviewer plays the customer. ^[8]Source 8Kong PS / Forward Deployed Engineer — Take-Home ExercisePublisheralexscottan (GitHub)Source typecandidate’s take-home repository The free case is that call: an AI customer with a building-permit problem, a 10-minute clock, and a score that quotes what you said. Run the free case before your take-home arrives, and practice asking the questions you’ll need for the build.

Questions people ask

What do forward deployed engineer take-homes ask for?

More than code, in the briefs we read. Tangible’s published FDE take-home asks for an ingestion service, a follow-up email to the client’s integration lead and an AI usage note, and Coframe’s asks for code plus a recorded video walkthrough.^{[1, 3]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany website

How long should I spend on an FDE take-home?

Keep to the time the brief states and say what you cut. Stated limits differ: Coframe’s published take-home gives 2 hours, and Ethyca’s says its challenge should take between 3 and 4 hours.^{[2, 3]}Source 2Ethyca Technical Challenge -- Forward Deployed Privacy Engineer (FDPE)PublisherEthyca (GitHub)Source typecompany websiteSource 3Coframe Forward Deploy — Take-HomePublisherCoframe (GitHub)Source typecompany website

Can I ask questions during an FDE take-home?

Some briefs invite it. Tangible’s published FDE take-home says to email questions during the assignment window, and AzurePartners’ says to email if something is ambiguous or to make a call and write it down.^{[1, 5]}Source 1Take-Home Assignment: Client Positions Feed OnboardingPublisherTangible (tangiblemarkets on GitHub)Source typecompany websiteSource 5Take-home: the planning agentPublisherAzurePartners (GitHub account)Source typecompany website

Can I use AI tools on an FDE take-home?

Follow the brief. Namastex’s published FDE take-home expects AI coding tools and asks for your AI conversations in an ai-logs folder, while Anthropic asks candidates to complete take-homes without Claude unless it says otherwise.^{[4, 15]}Source 4Desafio Técnico — FDE / AI Engineer (Namastex)PublisherNamastex (GitHub)Source typecompany websiteSource 15Guidance on Candidates' AI UsagePublisherAnthropicSource typecompany website

Keep reading

Guides

• The forward deployed engineer interview, round by round

Lessons

• The walking skeleton: a functioning first version
• Writing for customers: executive summaries and the follow-up email
• AI assistance rules, by company and by stage
• Sequence the work by risk and value

Questions

• Looking at your take-home submission, what would you change with another day, and what did you leave out on purpose?
• Present a past technical project to a mixed audience of engineers and executives, then take questions.
• How do you use AI coding tools in your daily work, and where do you not trust them?
• Write a script that scores model outputs against a ground-truth file and groups errors by type.

More from the blog

1. Interview rounds

   Prompt engineering interview tests: what they look like and how to work under the clock

   What timed prompt-engineering tests for FDE and applied AI roles look like, and a test-first way to work against the clock without guessing.

   10 min readPublished September 2026

2. Interview rounds

   The take-home walkthrough video: a script that shows your judgment in a few minutes

   Several FDE take-homes ask for a recorded walkthrough. A shot-by-shot script, a README template and the recording mistakes that hide your judgment.

   11 min readPublished September 2026

3. Interview rounds

   Voice AI agent take-homes: what companies ask for and what to get right

   Several FDE take-home briefs ask for a voice agent. What they ask for, and how to handle latency, interruptions, failed tool calls and hand-off to a human.

   10 min readPublished September 2026

More in Interview rounds, page 2