McKinsey Solve Game: Redrock, Sea Wolf, and the Score You Can't See

Solve arrives early, typically within a week of submitting your application, after the resume screen and before any human interview. You take it once, unproctored but rule-bound, on a PC or Mac. Results combine with your resume to decide who gets a first round. Strong resumes do not excuse weak Solve results, and a strong Solve does not rescue a resume McKinsey was not going to interview anyway. Both gates have to open.

How much it filters is not published. Coaching firms and forum self-reports converge on a pass-through around 20 to 30%, which makes Solve the single largest cut in the funnel. Treat those numbers as unofficial, but treat the conclusion as real: most applicants exit here, usually with less preparation than they gave their resume.

Plan for one attempt. Candidates report one sitting per application cycle, with another chance only after the standard reapplication wait, commonly 12 to 24 months. There is no published retake rule to appeal to. The practical reading: Solve is not a formality to click through the evening the email arrives. Schedule it inside the invitation window, after your preparation, not before it.

The standard invitation runs about 65 minutes and contains two games: Redrock Study, roughly 35 minutes of data analysis built around a wildlife research scenario, and Sea Wolf, roughly 30 minutes of constraint-based selection where you assemble microbe teams to cleanse ocean sites. Longer invitations, around 85 minutes, add a third module, most often the Sustainable Futures Lab, a judgment-and-tradeoffs scenario.

What you will probably not see: Ecosystem Building, the 8-species food chain game, retired from the default rotation between mid-2025 and early 2026 by prep-firm accounts. Plant Defense, the tower-defense module, left the rotation earlier. Both linger in a minority of invitations in specific regions and roles, which is why the invitation email matters more than any guide, this one included.

Read that email like a contract. It states your modules, your total time, and your window. If it says 65 minutes, prepare Redrock and Sea Wolf and skip legacy-game tutorials entirely; hours spent on Ecosystem strategy videos are hours spent on a game you will never play. If it names three modules or quotes a longer sitting, add the Sustainable Futures Lab notes below to your plan.

The product score is your answers: questions right in Redrock, sites cleansed in Sea Wolf. The process score is everything around them. Solve logs which data points you drag into the Research Journal, every entry in the on-screen calculator, your navigation between screens, how often you revise, how your pace varies across questions. McKinsey scores how you worked, not just what you produced.

This is the part of Solve you cannot cram, and it is the part most candidates never hear about. It exists because McKinsey is screening for how you would behave on an engagement: a consultant who checks the right data before concluding, works at a steady pace, and does not thrash. Erratic clicking, scattershot data collection, and answer-changing sprees all read as noise even when the final answers land correct. Careful and mostly right can outscore rushed and slightly more right.

You play for the process score with habits, not tricks. Read before you click. Collect data because a specific question needs it, not because it is on screen. Do arithmetic once, cleanly, in the logged calculator rather than seven exploratory times. Answer, sanity-check, move on. The telemetry cannot distinguish confident method from lucky method in one action, but over 65 minutes it absolutely can.

Phase one is investigation. You read the study, charts, tables, and methodology text, and drag data points into a Research Journal. Only a fraction of the numbers on screen matter, roughly a tenth by prep-firm counts, and your selections are themselves assessed. The journal is your only workspace later, so what you skip now is gone when you need it.

Phase two is analysis: a handful of calculation questions, typically three or four, answered with the on-screen calculator against the data you journaled. Percentages, growth rates, weighted averages, means and medians dominate.

Phase three is the report. You complete a pre-written summary by filling in computed values and choosing the chart, bar, line, or pie, that best shows the finding. The gate: once the report opens, you cannot return to investigation or analysis. It is a one-way door, and it is the mechanic that ruins sittings, because candidates discover a missing number with no way back to fetch it.

Phase four is a block of case questions, around six, on fresh mini-scenarios unrelated to your study data. Budget for them. Candidates who spend freely in the early phases hit this block with two minutes and guess through it.

Reported question mixes put calculation at 60 to 70% of Redrock. The recurring types: percentage change and percentage points, growth rates including compound growth, weighted averages, means and medians, simple probability, and ratio scaling. Nothing beyond confident high-school arithmetic. The bar is speed with zero slips, under logging.

The calculator changes the skill. You are not doing mental math; you are doing supervised calculator math, and the telemetry sees your entries. The practical standard: set up the calculation before touching the keys, run it once, believe it. A candidate who types 14 exploratory calculations for one answer looks like what they are, someone hunting for a number they cannot define.

Two traps recur. Percentage against percentage points: a wolf survival rate moving from 40% to 50% is a 10 percentage-point rise and a 25% increase, and Redrock asks in both currencies. And weighted averages: any question mixing groups of different sizes wants the weighted figure, and the unweighted mean is always sitting there as a wrong-answer option.

Drill this before the sitting: ten minutes daily of percentages, growth, and weighted averages, done on a plain calculator with the setup spoken aloud first. A week of that removes the slips that cost more than any strategy error.

After the report, Redrock switches format: around six standalone questions on mini-scenarios that share the wildlife theme but not your data. A mix of calculation, multiple choice, short reading-analysis, and choose-the-right-chart. Each is self-contained, so nothing from your journal helps and nothing you missed earlier hurts.

The ambush is structural, not intellectual. Candidates assume the report is the finale, spend down to it, and meet a third of Redrock's questions with the clock nearly dry. The questions themselves are the easiest in the module when given normal attention, which makes points lost here the cheapest points in the whole assessment.

Play it with a time budget set at the start of Redrock: roughly a third of the module held for everything after the report opens. Inside the block, triage. Calculation questions with clean setups first, since they score reliably. Chart-selection questions next, on one rule: the chart matches the claim, trend claims take lines, composition claims take pies, comparison claims take bars. Long reading questions last, because they price highest in seconds per point.

If the clock beats you anyway, answer everything. Reported scoring gives nothing for blanks, and a reasoned elimination guess under telemetry still reads better than an abandoned screen.

Collect against the report, not against the study. Redrock's report phase completes a pre-written summary, which means the questions are largely predetermined: population counts by period, rates of change, group comparisons, one headline conclusion. Read phase one asking "which numbers complete that summary," and the journal builds itself. Read it as "what is interesting here" and you will journal a tenth of the screen and still miss the number the report wants.

Concretely: on each chart or table, take the totals, the endpoints of any time series, and the figure for each named group. Skip narrative color, methodology defenses, and any number the text itself calls preliminary, unless a question flags it. When the methodology section defines a term, pack counts, survey coverage, sampling windows, journal the definition's numbers; definition questions are how Redrock tests careful reading.

Then stop. The journal is scored as selection quality, and over-collection is a visible behavior, not a safety blanket. A journal holding 12 relevant figures beats one holding 40 where the right 12 swim among decoys, both for the telemetry and for you in phase three, when you are searching your own journal under a closed door.

The rule that holds it together: every drag answers the question "what will I compute with this?" No answer, no drag.

You cleanse three ocean sites in about 30 minutes by assembling a team of three microbes per site. Each site defines seven characteristics: three numeric attributes with target ranges, and four binary traits, some desirable, some forbidden. A team succeeds when the three microbes' averaged attributes land inside the site's ranges, at least one microbe carries a desirable trait, and none carries a forbidden one.

The flow per site runs through set steps. You commit to two of the seven characteristics up front, a choice that shapes the microbe pool you will see. You then triage a batch of ten microbes, keeping some for this site, routing some toward the other site, rejecting the rest. You build out a shortlist through successive picks, and finally commit three microbes as the site's team. The first two sites include a confirmation step; the last runs slightly shorter.

Scoring is partial, reportedly about 20% per condition met, each attribute average in range, desirable trait present, forbidden trait absent. That matters strategically: a site can score 80% with one attribute missed, so a stuck site is a site to finish imperfectly, not a site to perfect while the third expires. Reported details of steps and weights shift between sittings; the averaging mechanic and the trait gates are the stable core to prepare.

Averages under constraints. Each attribute is scored on the mean of your three microbes, and means are forgiving: one microbe far outside a range is fine if the other two pull the average back in. That single fact drives the strategy. Do not hunt three microbes that each sit inside every range; that perfect trio rarely exists in the pool. Hunt combinations: a high, a low, and a middle whose average lands in the band.

Work the gates in the cheap order. Traits are binary and instant: a microbe carrying a forbidden trait is dead on arrival no matter how beautiful its numbers, so eliminate on forbidden traits first, note which survivors carry a desirable trait second, and only then do averaging arithmetic on what remains. Arithmetic is the expensive step; spend it on eligible microbes only.

For the averaging itself, use the target midpoint as an anchor. If the range is 20 to 40, you need the three values to sum near 90. Summing to a target is faster under pressure than computing means, and it makes compensation obvious: a microbe at 55 needs teammates around 20 and 15, and you can see that at a glance.

The trait conditions are reportedly worth 40% of a site between them, purchasable with no math at all. Bank them first, every site.

Commit to the constraints that are hardest to satisfy, because the choice shapes the microbe pool you draw from, and a pool shaped by your tightest constraint is a pool full of candidates that survive it. Committing to a constraint nearly every microbe satisfies anyway spends your influence on a problem you did not have.

Rank the seven at a glance. A forbidden trait is usually the sharpest filter: it kills microbes outright, so pointing the pool away from it saves the most later eliminations. Among the three numeric attributes, the tightest band relative to the values you see in the pool is the binding one; an attribute whose range covers most observed values is barely a constraint at all. The desirable trait needs only one carrier across your final three, which makes it the easiest condition on the board and almost never worth a commitment slot.

Then let the choice cascade into triage. Having committed to, say, the forbidden trait and the narrowest attribute band, sort the ten-microbe batch by exactly those two tests, and route the near-misses toward the site whose bands they do fit rather than the reject pile. The two-site structure rewards candidates who treat the batch as one shared inventory, and the telemetry watches whether your sorting follows any logic at all. Make it visibly follow this one.

The Sustainable Futures Lab, SFL, is the judgment module that appears mostly on the longer, roughly 85-minute invitations as a third game. Reported format: about 20 minutes, 13 questions, one opening drag-and-drop ranking and 12 scenario multiple-choice questions that follow a single sustainability-themed project as it hits complications. No math to speak of. It reportedly tests five things: prioritization, deciding under uncertainty, reading messy information, balancing trade-offs, and handling stakeholders.

Answer it like the project's manager, because that is the simulation. When the scenario forces a ranking, sequence by what de-risks the project: dependencies and blockers first, visible-but-cosmetic issues last. When options trade speed against quality or cost against relationships, favor the choice that protects the critical path and keeps stakeholders informed before it escalates, purely speed-maximizing and purely conflict-avoiding options are both scored traps. When information is incomplete, prefer answers that act on what is known while naming what would change the decision, over answers that stall for certainty.

Consistency matters more than any single question. The module presents one evolving storyline, and a candidate who prioritizes delivery in question 3 and abandons it for harmony in question 9 reads as unprincipled rather than flexible. Pick the operating logic above and apply it all 13 times.

Rarely, and only if your invitation says so. Ecosystem Building, the 8-species food chain game that defined Solve for years, left the default rotation between mid-2025 and early 2026 by prep-firm accounts. Plant Defense, the tower-defense module, was phased out earlier. Regional and role-specific sittings occasionally still carry legacy modules, which is the only reason to check for them at all.

If your invitation does name Ecosystem, the game in brief: you build a self-sustaining food chain of eight species from a larger pool, on terrain whose conditions, depth or elevation, temperature and the like, each species must tolerate. Survival is calorie accounting: every species' calories needed must be covered by what it eats, without exhausting the calories its prey provide. The classic method still works, pick the terrain first, anchor on producers, and verify the calorie ledger from the bottom of the chain up before submitting.

For everyone else, the actionable point is negative: skip legacy-game preparation entirely. Most Solve content online was written for Ecosystem and Plant Defense, and it goes stale slower than the assessment changes. An hour on an Ecosystem walkthrough video is an hour not spent on Redrock percentage drills or Sea Wolf averaging, for a game your invitation never mentioned. Let the email decide, then prepare only what it names.

Day one: logistics and rules. Confirm your modules and timing in the invitation, run the tech check on the machine you will use, read McKinsey's official instructions, and block a quiet 90-minute window inside the deadline, not on it. Solve is PC or Mac only, single sitting, no external tools.

Days two through six, three daily drills. First, ten minutes of Redrock math on a plain calculator: percentage change, percentage points, compound growth, weighted averages, setup spoken aloud before each. Second, one data-extraction rep: take any dense chart or table, an annual report page works, and pull the totals, endpoints, and per-group figures in under three minutes, the Research Journal motion. Third, one Sea Wolf-style averaging drill: three numbers summing into a target band, chosen from a list under a timer, plus a pass of eliminate-by-trait logic. Fifteen to twenty minutes total; consistency beats volume because you are training defaults the telemetry will watch.

Day six, one full timed simulation if you are using a third-party simulator, treated as a dress rehearsal: same machine, same desk, no pauses. Day seven, nothing but the sitting itself, rested. The process score reads like a fatigue detector, steady pacing, clean selections, no thrash, and sleep moves it more than a final cram session does. Take Solve the way you would run an engagement week: prepared early, executed calmly.

Substantially, in both directions, which is good news because Solve is the doorway to the case interviews, not a substitute for them. Redrock's math is case math with a calculator: the same percentages, growth rates, and weighted averages, the same discipline of defining the setup before computing, the same so-what standard of tying a number to a conclusion for the report. Candidates who have drilled McKinsey case interview questions walk into Redrock with the hard part done. Sea Wolf's constraint triage, find the binding constraint, spend effort only on eligible options, is structuring logic wearing a costume.

What does not transfer: everything verbal. Solve gives no credit for reasoning aloud, no partial credit for a well-framed wrong answer, no interviewer to read and adjust to. Candidates who lean on communication polish find Solve indifferent to it. The compensating skill is interface discipline, working cleanly under logging, which case prep never teaches and the drills above do.

And the reverse direction is the schedule warning. Passing Solve triggers the live rounds, sometimes within a couple of weeks: interviewer-led cases plus the PEI, the personal experience interview, in every session. If you start case and PEI preparation only after your Solve result arrives, you have given yourself days for the part of the process that needs weeks. Run them in parallel from the day you apply.