Where we start
Ren, the constructed analyst moving into a first data scientist role, has done the heavy lifting: a job description read into eight competencies, a bank of eight stories, a SQL habit of reading the question before typing, practice against the bench, and a way through the statistics and product case rounds. What is left is the part of the loop that happens around the rounds. A constructed take-home brief has arrived with "expect about 3 hours" at the bottom. A panel will end with "any questions for us?". And there will be an evening after every round with nothing scheduled in it. Tonight those three become habits rather than accidents.
★ What you walk out with tonight. A take-home read into the competencies it checks, a clarifying email ready to send, a time-box plan whose hours add up, a README skeleton you can reuse on any take-home, three questions for your panel that tell you something you need to know, and a five-line review template you fill within the hour of every round, with Ren's worked example beside it.
A take-home is a work sample with no one watching the clock
5 min liveA take-home sends you a brief and some data and asks for an analysis back. It is the round closest to the job itself, which is both why it is useful and why it eats evenings. Two habits change it: decide the hours before you open the data, and write the README before you write the analysis. The reviewer reads your write-up first and your code second, so the write-up is the biggest block of the time-box, not whatever is left over.
LiveWhat the evidence does and does not say3 min▶
No peer-reviewed study of take-home assignments' validity was found. The searches turned up vendor blogs and forum threads, not research. The nearest evidence is the literature on work sample tests, where a candidate does a piece of the actual job. Schmidt and Hunter (1998) put work samples at .54 as a correlation with later rated job performance; Sackett, Zhang, Berry and Lievens (2022) revised that to .33, one of the larger drops in their table (session 1 has the argument about why the numbers moved).
The link to take-homes is an analogy, and it should be said as one. The work samples in those studies were supervised and built for a specific job; a take-home is done alone, at home, in however many hours the candidate chooses to spend. That difference is exactly where the practice below comes from: if the hours are not controlled by the room, you have to control them yourself.
LiveAsk the hours, keep the hours, read the brief3 min▶
Ask how long they expect it to take, and keep to it. If the brief does not say, ask. If it says "about 3 hours", plan for 3 and stop near 3. Say in the README where you stopped and what you would do next. A reviewer comparing candidates cannot see the hours, but can usually tell when one submission had a weekend and the brief asked for an evening, and some will read that as not following the brief.
Read the brief into the competencies it is checking. The same move as session 1, on a smaller text. "A CSV of orders" checks data quality and wrangling. "Recommend whether to" checks whether you can turn a vague ask into a question and explain a result to someone who will act on it. The time limit checks whether you can scope. Each competency gets a line in the README.
Self-studyShow your checks, skip the gold plate3 min▶
Show the checks, in the README, as numbers. Row counts before and after each join, so the reader sees nothing duplicated or dropped (session 3's grain check, written down). NULLs counted in every column the analysis uses. One sanity check against something you can know independently, such as total orders matching the row count of the raw file. A check that was run but not shown does not exist for the reviewer.
Do not gold-plate. A dashboard, a model with tuned hyperparameters and a slide deck were not asked for. Extra polish spends the hours that should have gone into the write-up, and it can bury the one recommendation the brief wanted under three things it did not. If you see a bigger idea, it goes in the fifth README line, "what I would do with more time", as one sentence.
The beautiful notebook with no answer in it. A pattern reviewers describe often: a candidate submits a long notebook with clean charts, a feature engineering section and a model comparison, and the reviewer scrolls for some time before finding that the question the brief asked, should we do this or not, is never answered in a sentence. The work was real. The README that would have put the answer at the top was missing, and the reviewer had other submissions to read.
Ask to learn whether you want the job
5 min liveIn a structured interview your questions are held until after the interview proper, and they are usually not part of the score. That changes what they are for. They are not a final test to pass with a clever line. They are your one scheduled chance to measure the job the way the job has been measuring you, and the answers are data for a decision you will have to make.
LiveWhere your questions sit in a structured interview3 min▶
Campion, Palmer and Campion's 1997 review listed fifteen components of interview structure. The seventh content component, as listed by Levashina, Hartwell, Morgeson and Campion (2014), is no questions from the applicant until after the interview. The reason is measurement: questions from the candidate in the middle change the conversation from one candidate to the next, which is the inconsistency structure is trying to remove. So in a structured loop the interviewer finishes the rated questions, puts the pen down, and then asks what you would like to know.
What follows from that is modest and useful. Your questions are usually not rated, so a sharp question will not rescue a thin answer, and an ordinary question will not sink a strong one. No peer-reviewed effect size for "good questions" was found, so any claim that a particular question "impresses" interviewers is folklore. The questions are yours to spend on information.
LiveThree families, one question from each3 min▶
The work itself and its data. What you would work on first, what the data looks like, who owns its quality. The answer tells you whether the job matches the description you prepared for, and whether "data scientist" means models, dashboards or cleaning other people's tables. How decisions get made. Who uses the analysis, what happened the last time an analysis disagreed with a plan. The answer tells you whether your work will change anything. How success is measured in the first months. What good would look like after three months, how the role is reviewed. The answer tells you what you will be judged on once the interviews stop.
Ask different interviewers different families. The hiring manager can answer the third best; a future peer, the first; someone from the team that uses the analysis, the second. Write the answers down in your review afterwards: they are evidence too.
Self-studyAsk the recruiter what each round covers2 min▶
The most useful question of the whole loop is asked before it starts: what does each round focus on? Recruiters expect it, and the answer corrects the map you drew from the job description in session 1. Ask it again before the take-home: how long is it expected to take, and how will it be reviewed? Neither question is scored either. Both change how you spend your hours.
The question that changed the decision. A familiar story from people who have moved roles: a candidate asks, almost as an afterthought, "who reads the analysis, and what did they do with the last one?" and hears a long pause, then an honest answer that most reports go into a shared folder. That one answer told the candidate more about the job than the three rounds before it. Whatever they decided, they decided it with the information in hand.
The hour after the round is part of the prep
5 min liveMemory of an interview starts editing itself almost at once: the bad answer grows, the good ones blur, and by the next morning you remember how it felt rather than what was asked. Writing five lines within the hour turns each round into data for the next one. Over a loop of four or five rounds, those notes are the difference between practising in general and practising the exact gaps that showed up.
LiveFive lines, within the hour3 min▶
Write them on your phone on the way out if you have to. 1 What was asked, as close to the words as you can. 2 Which competency it was probing, named from your session 1 list. 3 What you answered: which story or which query, in a line. 4 What evidence was missing: a result you did not state, a check you did not say out loud, a NULL you did not mention. 5 One change for the next round, only one.
Line 2 is what makes the rest useful. "That went badly" cannot be practised. "The disagreement question, and my story had no result" can be practised tonight. The structured interviewer was scoring per competency; your review uses the same grain.
Self-studyFeed the bank and the drill list3 min▶
A review that stays in a notes app changes nothing. Line 4 goes somewhere. A behavioural gap goes into the session 2 story bank: add the missing result to the card, retag the story if it evidenced a different competency than you thought, or mark a competency thin if no story fitted. A technical gap goes onto the SQL drill list from session 3, and you run that trap again on the session 4 bench until it fails for the right reason when you plant it and passes when you fix it.
The answers to your own questions go in too. If the hiring manager said good at month 3 means "owns the weekly experiment readout", that is a competency to check your bank against before the next round.
Self-studyRen's loop, closed (constructed)3 min▶
Ren is constructed, and so is this loop. Across a recruiter screen, a SQL round, a take-home, a behavioural round and a panel, Ren filled five reviews. The screen review corrected the round map: there was no separate statistics round, so experiments would come up in the panel. The SQL review named one trap, a WHERE on the right-hand table of a LEFT JOIN, and it went onto the drill list and back through the bench twice. The take-home review noted 3 hours 10 minutes against a 3-hour brief and a README whose fifth line did the job it was meant to. The behavioural review is the one worked in step 6 below. The panel review recorded the answers to Ren's three questions.
What happens next for Ren is not on this page, on purpose. Nothing here claims an outcome, and nothing describes how any real employer runs a loop. What the routine buys is narrower and dependable: every round leaves Ren better prepared for the next one than the hours alone would have.
The same gap, three rounds running. A pattern career coaches describe: a candidate leaves four interviews in a fortnight each feeling "it was fine", writes nothing down, and only when someone asks them to list what they were asked do they notice that three rounds asked about influencing a stakeholder and they told the same story without a result each time. Five lines after the first round would have caught it two rounds earlier.
Ren's take-home, questions and review, then yours
★ 22 min · everyone buildsEvery artefact below is something you will use in a real loop. Ren's take-home brief is constructed: a CSV of orders from a constructed online shop, a request to recommend whether to change the free-delivery threshold, and "expect about 3 hours". Work each step on Ren's brief first, then write your own version for the loop you are in. Your numbers will differ.
Read the brief into competencies. One line per phrase, the competency it checks, and where it will show in your submission. If a phrase hides a choice the brief did not make, mark it: that is step 2.
Write the clarifying email before opening the data. Two or three questions at most, each one a choice that would change the analysis. End with the assumption you will make if nobody answers, so silence does not block you.
Plan the time-box, and check that it adds up. Book the write-up first, then the checks, then whatever is left for analysis. Keep a small buffer. Ren's hours are constructed; yours will differ.
Write the README skeleton first. Five headings and a placeholder under each, before any analysis. Fill the first two now; the rest fill as you go. Reuse the skeleton for every take-home.
Write your three questions for the panel, one from each family. Check each one: would the answer change whether you want the job? If not, swap it.
Fill an after-round review, then update the bank. Ren's is for a constructed behavioural round. Write yours within the hour of your next real round, and make line 4 land somewhere before you close the note.
Keep the five templates in one place. Brief reading, clarifying email, time-box, README skeleton, questions and review all fit in one document. The second take-home and the third review take minutes instead of an evening, because the thinking is already written down.
Before your next round
40 min- Write your README skeleton with the five headings and the three checks, and save it where you will find it when a take-home arrives.
- Write three questions for your next panel, one per family, and test each against "would the answer change my decision?".
- Make a review template with the five lines and a "bank update" line, on your phone, ready to fill on the way out of a round.
- Review one past interview from memory. Fill the five lines for it, even if it was months ago, and put whatever line 4 says into your story bank or SQL drill list.
Three questions
Clear all three and session 6 gets stamped in your passport.
1 · A take-home brief says "expect about 3 hours". Which plan follows Part 1?
No peer-reviewed study of take-home validity was found, so the practice rests on the work-sample analogy and on the brief itself. Keeping to the stated hours, putting the question and result first, and showing the checks answers what was asked. A and D gold-plate and spend the hours the write-up needed; C hides the answer from the reader.
2 · Why are the questions you ask usually not part of the score in a structured interview?
The component exists to keep the measurement consistent across candidates. No peer-reviewed effect size for candidate questions was found, which rules out B. The questions are still worth asking, which rules out D, and C misreads the component: questions are held to the end, not refused.
3 · Ren leaves a behavioural round that "felt fine". What does the review routine ask for?
Naming the gap as a competency is what makes it practisable, and the bank or drill list is where it turns into the next round's answer. A loses the details to memory, B records a feeling rather than a gap, and C spends the hour without writing anything down.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 6 cheat sheet · pin this
This is the last session, and the loop it closes is yours to keep running. Go back to the course home and the knowledge map to see how the six sessions connect, and reread session 1's prep plan with five reviews' worth of evidence beside it. For the other side of the table, structured interviews as the interviewer runs them, see learn-ai-hr ↗. For the reputation and portfolio that arrive before you do, see learn-personal-branding ↗.