A checklist for take-home assignments that get callbacks
I have reviewed take-homes on a phone, in a parking lot, between meetings. Assume ten minutes of skimming from someone who is tired. Almost every rule below follows from that one fact.
By Lin S., founder and coach at Open Loop
Here is the uncomfortable truth about take-home review. Your assignment took you six hours. It will get ten minutes, maybe fifteen if the first screen is interesting, from a person who has four of these to get through before Thursday and who is doing it on top of their actual job.
I am not defending this. It is just what happens, at every company I have worked at, and once you accept it the rules mostly write themselves. The candidates who do well are not the ones who did the most work. They are the ones whose work could be understood quickly.
A brilliant analysis nobody can find is worth less than an adequate one on the first page.
Put the answer first
The most common structural mistake is the chronological notebook. Load data, inspect nulls, plot twelve distributions, engineer features, train three models, and then, in cell 47, a conclusion. That is the order you did the work in. It is a terrible order to read it in.
What most people submit
Customer Churn Analysis. Section 1: Data Loading and Cleaning. “First, let's load the dataset and examine its structure...”
- Twenty minutes of reading before the reviewer learns anything
- The recommendation is at the bottom, if it exists at all
- Reads like a lab notebook rather than a deliverable
What gets a callback
Churn Analysis: Recommendation. Target the 30-to-60 day window with a save offer. 62% of churn happens there, it is concentrated in customers who never used the mobile app, and a 5-point reduction is worth roughly $1.4M a year. Confidence: moderate. Details and caveats below.
- The reviewer can stop reading here and still know your answer
- A number, a mechanism, and a size
- States its own confidence, which builds trust rather than eroding it
Notice the last line. Volunteering that your confidence is moderate does not weaken the submission, it strengthens it, because the alternative is a reviewer discovering the caveat themselves and wondering what else you were quiet about.
The structure that works
- 01
Recommendation
1 paragraph
What you would do, how big it is, and how confident you are. Written for someone who will read only this.
- 02
How you got there
3 to 5 bullets
The path, not the process. “Segmented by tenure, found churn concentrated in month two, checked whether it was a mix effect, it was not.”
- 03
Assumptions and what I did not do
1 short list
The highest-value section in the whole document and the one most people omit entirely. More on this below.
- 04
Evidence
3 to 5 exhibits
Your best few charts, each with a title that states the finding rather than naming the axes.
- 05
With another week
3 bullets
What you would do next and why it matters. This is where you show taste, which is what senior reviewers are reading for.
- 06
Appendix
as long as needed
Full exploration, model comparisons, the code. Clearly labeled as optional reading.
The assumptions section is the cheat code
Every take-home dataset is ambiguous somewhere. Dates that do not line up, a column with no documentation, a definition that could go two ways. You have to make a call, and candidates handle this in one of three ways.
- Silently pick one and hope nobody notices. Most common, and it reads as either careless or unaware.
- Email the recruiter with eight questions. Occasionally right, usually reads as someone who cannot proceed without permission.
- Pick one, write it down, say why, and note what would change if the other reading were correct. This is the answer.
If there is a model, there is a baseline
This is the fastest way to lose a modeling take-home. You submit a gradient boosted model with an AUC of 0.84 and no baseline, and the reviewer has no idea whether that is good. It might be worse than predicting the majority class.
Always report the dumbest thing that could work. Predict the base rate. Predict last month's value. Use a single-feature logistic regression. Then your real model has something to beat, and the comparison is the finding rather than the number.
While you are there, say out loud whether you checked for leakage. Take-home datasets are frequently seeded with a column that is a proxy for the target, sometimes on purpose. A suspiciously strong model that nobody interrogated is a worse outcome than a mediocre model with a paragraph explaining why you dropped cancellation_reason from the feature set.
What the reviewer is actually asking
| The silent question | What answers it |
|---|---|
| Did they answer the question I asked? | A recommendation in the first paragraph that maps to the prompt's wording |
| Would I repeat this number in a meeting? | Stated assumptions, a sanity check, and a caveat you volunteered |
| Can I read their code in six months? | Named steps, no dead cells, no commented-out experiments, runs top to bottom |
| Do they know what they did not do? | A with another week section that names real tradeoffs |
| Do I want to review this person's work weekly? | Brevity, honestly. This is scored even where nobody admits it. |
Include and cut
Include
- A README with how to run it and how long it takes
- Chart titles that state the finding: “Churn concentrates in days 30 to 60”, not
Churn by tenure - One sanity check you ran and its result, even a boring one
- A stated time budget: “spent about 4 hours, prioritized X over Y”
- Pinned dependencies, or at least a requirements file
- The question you would ask the business before doing this for real
Cut
- A correlation heatmap of forty columns that you never reference again
- Every distribution plot you looked at while orienting yourself
- Three models with no baseline and no reason for choosing between them
- Commented-out cells,
df2,df_final,df_final_v2 - An apology in the introduction about what you ran out of time for
- Any chart you cannot say the point of in one sentence
How long to actually spend
If it says four hours, spend somewhere between four and six, and say what you spent. Do not spend twenty. Two reasons, and the second one is the one people miss.
The obvious reason is your time, which in a live search is your scarcest asset. The one candidates do not think about is that a submission which clearly took twenty hours is itself a negative signal. If you cannot triage a scoped four-hour task, that tells the reviewer something about how you will handle a quarter. I have seen genuinely excellent twenty-hour submissions get a soft no with the note “did not respect the time box”, and the candidate had no idea that was the reason.
Spend the time in roughly this shape: half on understanding the data and the question, a quarter on the analysis, and a quarter on the writeup. That last quarter feels wasteful and it is the part that determines the outcome.
Assume you will present it
In most loops the take-home is not the end. It becomes the artifact for a follow-up round where two or three people ask you about it, and that round is where the decision really gets made. Prepare for three specific questions, because you will get some version of all of them.
- 01What would you do differently with more time? Have three real answers ranked by value.
More feature engineeringis not one of them. - 02What is the weakest part of this? Answer honestly and specifically. Candidates who claim there is no weak part fail this question completely, and there is always a weak part because you had four hours.
- 03Walk me through this number. Pick your headline number and be able to trace it from raw data to output without opening the notebook. If you cannot, you do not know your own analysis well enough.
One habit that pays for itself: keep a scratch file while you work, and write down every decision you make and why, in the moment. It takes seconds and it means that when someone asks in a week why you filtered those 900 rows, you have the answer instead of a plausible reconstruction. Reviewers can tell the difference between the two more often than you would think.