Case Interview Drills 2026: Score One Skill at a Time

What one scored case interview drill contains

The runner is one file, src/components/tools/drill-runner.tsx, and it does five things in a fixed order. Reading the code is the only way to know what a drill returns, so that is what this page does, checked 29 September 2026.

1. It shows the prompt, plus any numeric context as separate tiles. The margin drill shows four: volume, price, cost per unit, fixed costs.
2. It starts the clock on your first keystroke. Opening the page costs nothing, which matters because the timer is there to imitate the moment you start speaking.
3. It scores on submit, locally. There is no network call in the runner and no browser storage, so your answer never leaves the tab.
4. It renders the scorecard: the total, then each dimension as a bar with its own points out of its own maximum, then a block called "Fix first", then a block called "Kept".
5. It shows a model answer written for that prompt, and the next button cycles through the list.

Two rules decide the two feedback blocks, and both live in src/lib/grading/rubrics.ts. A dimension enters "Fix first" when it scores below 60 percent of its maximum. It enters "Kept" when it scores the maximum exactly. Nothing in between is commented, which keeps a debrief to two or three lines rather than four paragraphs nobody reads.

The stock of prompts is small and worth stating plainly. The repository holds eight open-ended drills: two math, two structure, one brainstorming, one market sizing, one exhibit, one synthesis. The separate math practice page uses a generator instead, which builds a set of 30 items from eight calculation types, four items each for the first six types and three each for the last two.

The six rubrics, and who wrote the weights

Every weight below is read from our own code. We wrote them. No consulting firm publishes a case interview rubric, and it is worth being blunt about that, because a table of weights invites the assumption that a firm handed them over.

Each set totals 100 points. The reasoning for each weight sits on how an answer is assessed, a page that runs the six rubrics live on a sample answer and prints what they return, so it cannot drift from the code that scores you.

What firms do publish is the list of moves they expect, and that is what the six types split apart. BCG describes the task on its case interview preparation page as working through a problem step by step: "structure your approach, ask thoughtful questions, analyze data, perform quick calculations, and identify the most important factors." Bain asks candidates on its own case interview page to "share your reasoning and calculations, even if you're unsure." Reading their sentences against our dimensions is a reading, and not an equivalence: Bain asking for visible calculation is the reason Working carries 25 points, but Bain never named a number.

One margin drill from the case interview set, scored three ways

Here is the stock math prompt. A retailer sells 240,000 units a year at $45. Cost of goods is $27 per unit and fixed costs are $2.4M. What is the annual operating profit?

The arithmetic runs in three steps. Contribution per unit is $45 minus $27, so $18. Across 240,000 units that is $18 times 240,000, which is $4,320,000. Subtract $2.4M of fixed costs and operating profit is $1,920,000.

Three answers, all submitted to the real rubric on 29 September 2026, came back like this.

The gap between the first and third row is the whole argument for scoring dimensions separately. All three rows contain the right number. The first would be marked correct by any answer key and would still lose half the available points in a room, because the candidate did nothing with it.

Read the second row again, though. Units scored 4 out of 15 because the answer used bare digits. That is a real interview habit, and the rubric caught it on a reply a human grader would probably wave through.

Three scores we can measure and would not defend

Every scored tool has a soft spot, and ours can be measured rather than guessed at, because the scoring is deterministic. The same answer returns the same score forever. Here are three results we got on purpose.

1. Three keywords score 85. Submitting the string revenue; cost; competitor to the structure drill returns 85 out of 100. Coverage gives 45 out of 45, because it looks for any synonym of each expected branch. Separation gives 25 out of 25, because semicolons read as a list. Only Depth drops, to 5 out of 20. Nobody would pass a real structure round with three words.
2. A wrong number keeps 72. Answer the margin drill with a well-written derivation that lands on $4.32M rather than $1.92M, and the score is 72. Answer falls to 12 out of 40, while Working, Units and Interpretation stay at maximum, because those three dimensions judge form.
3. A long opening sentence beats the answer-first test. Synthesis checks the first sentence only. Put the call at the end of one long opening sentence and Answer first pays 35 out of 35. Put the same call in the final sentence and it pays 0.

None of this is a bug to be quietly patched, and the third one is close to how an interviewer listens. It is the price of a scorer that judges the presence of the expected moves rather than the truth of a claim. Bain says on the page cited above that it is "looking for how you think" rather than one right answer. A rubric can see the shape of your thinking. It cannot see whether your premise is false, and a confident answer built on a false premise scores well here and dies in the room. No language model is called anywhere in that path, and a grader built from regular expressions and counts is what makes soft spots like these three measurable at all.

Running six case interview drills in forty minutes

The routine below is the one the weights imply, and it takes about forty minutes end to end. Timings are ours, from running the set.

1. Math first, four minutes with the clock visible. Write the calculation on the page rather than in your head, because Working and Units are 40 of the 100 points and both are invisible in a mental answer.
2. Structure next, five minutes. Lay out branches as separate lines and push one level under each, since Depth is the dimension that three headline words cannot fake.
3. Market sizing, six minutes. Open with what you include and exclude, then chain at least three assumptions, then close with a plausibility check against a population or a GDP.
4. Exhibit, five minutes. Quote one figure, name a cause, then say what the client should do. So what carries 30 points and is scored on that last move alone.
5. Brainstorming, four minutes. Aim at six ideas before stopping, group them, then name the one you would start with.
6. Synthesis last, three minutes. Recommendation in the opening sentence, the number behind it, the risk that would reverse it, the first step.

Then read your six "Fix first" lines together. If the same dimension appears three times, that is your week, and repeating the other five is comfortable rather than useful. Drills are how you find that dimension. Whole cases are how you test it under a conversation, and McKinsey publishes four complete ones with suggested answers on its interviewing page, free, which is a better next step than a seventh drill.

What a case interview drill here will not do

It does not listen. There is no voice input in the runner, so an answer typed at your own pace is easier than the same answer spoken under a stare. It does not save. No account, no progress, no history: close the tab and the session is gone, which is a deliberate consequence of scoring in the browser.

It does not adapt. Difficulty is fixed per prompt, there is no ladder, and the next button loops back to the first drill once you reach the end of a list of two.

It does not judge whether you are right about the business. The scorer looks for a number, a list, a cause, a risk, a next step. It cannot tell a sensible assumption from an absurd one, and the comment in the code says exactly that.

And it is not an assessment simulator. Our assessments page says "0 of 6 simulators are live" and the count is honest: the case content and timing rules are extracted, the interactive runs are not built. If you came looking for a Casey or a Solve replica, the drills are not it. They train the skills those tests read, which is a smaller claim.

Opening one, and what the paid side costs

Start with the structure drill if you do not know which dimension you lose, because Coverage and Depth are the two that separate fastest. Every drill page opens with no account and no email. The six linked from our home page are the six rubric types, and the calculators sit on the tools index, where 4 of the 7 listed entries are live today.

The paid side, plainly. A subscription runs $49 a month, $120 once for a 90-day sprint, or $249 a year, and packs exist at $29 for 50 drills, which is $0.58 each, and $20 for five full cases. What money buys is volume and the guided case runs. It does not buy a better grader, because the grader on the free drills is the same code, and you can read its weights before you pay anything.

One honest flag from the same reading. Our plan cards list a 50-drill pack, while this repository holds the eight open-ended prompts and the 30-item math generator described above. Treat the eight and the thirty as what you can verify today, and the drill pack count as a production claim we have not checked line by line.