The Presentation Capstone: Analysis to Storyline to Deck

Presentation track, branch pr-capstone (Capstone) · Node PR7.01 · ~14 focused hours · CROWN · Mastery gate ≥ 90%

Food revenue grew 57.9% last year. Pharma revenue grew 16.9%. A board that reads only those two numbers puts its next ₹12 crore into food, and the decision is wrong by 10.6 percentage points of annual return.

Both numbers are true. Both come off the same page of the same management pack. The reason they mislead is that revenue growth measures how much space you filled, and the board is deciding what space to build, which is a question about the return on a pallet slot rather than about the revenue that passed through one. A capstone submission that leads with the growth rates will be aligned, sorted, correctly coloured and confidently delivered. It will also lose the company about ₹1.27 crore a year, every year, for as long as the capacity stands.

That gap is the thing this branch exists to close, and it is why the capstone is scored twice. Seven machine rules can tell you that a slide is clean. Nothing in the seven can tell you that the argument on it is the argument the room needed. A deck built on the food-growth story scored 0.91 on the machine and 0.55 on the rubric, which blends to 0.766 against a crown gate of 0.90. A deck built on the capacity-return story scored 0.9389 on the machine and 0.90 on the rubric, blending to 0.9233. The machine barely separated them. The rubric separated them by 35 points, and the rubric is the part that mirrors what a real board would do.

The capstone runs on a clock. One hundred and fifty minutes, a brief you have not seen, a dataset nobody has cleaned for you, and a stated audience with a decision pending. Eighty-five of those minutes go before a single slide exists, which is 56.7% of the budget, and the candidates who miss the gate almost always miss it by inverting that ratio. They open the software at minute twenty because a blank slide feels like progress. By minute ninety they have eleven attractive exhibits and no idea which four of them carry an argument, and there is no longer time to find out.

Two reference solutions run through the pages below, each worked from a brief to a scored deck. One is an Indian cold-chain operator deciding where to put growth capital. One is an American freight carrier deciding between a price increase and a capital program. Each contains a wrong turn that the analyst takes for real and then corrects, because a reference solution that goes straight from brief to answer teaches the wrong thing about how analysis actually feels at minute forty.

Nothing new is taught here. The message sequence comes from the foundations branch, the mechanics from the second, the chart grammar from the third, the storyline from the fifth. What is new is the compression: doing all of it at once, on unfamiliar material, with a clock running and no one to ask.

Learning objectives

You can:

1. Read a capstone brief and extract its four fixed points (the audience, the decision, the medium and the time budget) in under ten minutes, and state the one sentence the whole submission has to support.
2. Run a six-check sweep on an unfamiliar dataset in twenty-five minutes: shape, ties, coverage, definitions, impossible values, and the question the data cannot answer, producing a written list of what the analysis may and may not claim.
3. Catch a false claim in the brief itself, by recomputing the figure the cover note asserts and reporting the difference in both points and relative terms.
4. Compute the decision number rather than the headline number, distinguishing a growth rate from a return on the resource the decision actually allocates.
5. Build a ghost deck against a page budget derived from the meeting length, and make the cut on paper, recording each cut page with the reason it was cut.
6. Select an exhibit set of four from a candidate list of ten or eleven, stating for every rejected candidate the message it would have carried and why that message does not advance the decision.
7. Write a title sequence that survives the read-only-the-titles test, with every title a full sentence carrying the number that makes it checkable.
8. Pick the archetype the audience expects and defend the choice against the one obvious alternative, including what changes when the medium changes from a pre-read to a presented deck.
9. Run the seven-rule self-review in the order that finds the most defects, score each slide, and compute the deck composite as 0.6 machine plus 0.4 rubric.
10. Produce the submission set (deck, storyboard, one-page rationale, delivery notes) and defend every judgment call in it, including the one you got wrong first.

Prerequisites & connections

Builds on. Everything in the eight branches, as skill rather than as content. The foundations branch supplies the takeaway-audience-decision-medium sequence you will run in the first ten minutes. The mechanics branch supplies the master, the grid, the type scale and the object discipline that make a nine-page deck consistent without a second pass. The chart-craft branch supplies the form heuristic, the declutter order and the action title. The visual branch supplies the ghost deck and the diagram families. The storytelling branch supplies the pyramid, the horizontal and vertical logic and the arc. The archetypes branch supplies the six deck shapes and their templates. The delivery branch supplies the ninety-second opening and the question taxonomy. None of these is a hard gate. A learner who has completed only the chart branch can attempt the capstone, and will find out quickly which branch to go back to.

Feeds forward. Nothing inside the Presentation track. The capstone is the last node in the galaxy, and what it feeds is the work you do afterwards: the finance capstone's communication rubric, the consulting engagement read-out, and the research deck you will actually be paid for. Learners who hold both this crown and the finance galaxy's capstone should present one artefact to both rubrics and note where the two disagree, because the disagreements are informative and there are usually two of them.

The deliberate non-overlaps matter more here than anywhere else in the galaxy, because a capstone that re-teaches its prerequisites is a review session wearing a crown. PR0.01 owns the audience profile, the medium choice priced in reader-minutes and the three arrival failure modes; the capstone applies that sequence under a clock and does not re-derive it. PR2.01 owns chart-form choice and the shared form heuristic; the exhibit-selection section below assumes you can already name the right form and asks a different question, which is whether the exhibit earns a page. PR2.03 owns the single chart's action title, its focal point and its layout; the capstone owns the title as a sequence and hands the individual title back. PR3.02 owns the ghost deck as a practice; the capstone owns the ghost deck against a clock and a page budget. PR4.01 and PR4.02 own the pyramid and the horizontal and vertical logic of titles; the capstone runs the read-only-the-titles test on its own output and cites the rule rather than teaching it. PR4.03 owns the narrative arc and the big idea. PR5.01 to PR5.03 own the archetypes and ship the templates the capstone applies. PR6.01 owns rehearsal, Q&A and remote delivery; the delivery notes here are a submission artefact, not a delivery lesson.

One more boundary, and it is a hard one. The capstone brief and its dataset appear nowhere else in the galaxy. Neither Tarangini Cold Chain nor Harborline Freight Systems is used in any other Presentation module, and neither is a real company. That is the same rule the finance and consulting capstones follow, for the same reason: a capstone that reuses a dataset the learner has already seen measures recall rather than judgment.

Core teaching content

1. The brief, the clock and the rules of engagement

A capstone brief is one page and contains four fixed points. Everything else in it is context you may use or ignore.

The audience is a named role with a stated number of people, not a category. "The board of Tarangini Cold Chain: four directors, of whom one represents a minority investor" is an audience. "Management" is not. The decision is stated as something that gets settled in the meeting: allocate ₹12 crore of FY27 growth capital between pharma capacity and food capacity, or hold it. The medium and slot are given: a nine-page pre-read circulated 48 hours ahead, walked in a 25-minute board slot. The time budget is the clock on your own work: 150 minutes from opening the brief to submitting.

Those four points are fixed because a capstone that lets you choose your own audience lets you choose an audience your analysis happens to suit. Real briefs are constrained in exactly this way, and the constraint is most of the difficulty.

Three rules of engagement govern the attempt, and each exists because of a specific way candidates fail.

You may not extend the clock. A submission at 165 minutes is graded and then discounted, because the whole point of the exercise is the triage that happens when time is short. Candidates who run over are almost never running over on the build; they are running over because they did not cut on paper at minute seventy-five.

You may not add data. The dataset is what the board has, and a recommendation resting on a number nobody in the room can see is a recommendation nobody in the room can approve. Where you need a figure the data does not contain, you say so in the rationale, name what you would have asked for, and state how the answer would change if it came back the other way. That is a scored line, not a penalty.

You may not use an assistant to write the takeaway, choose the exhibits or draft the titles. Compression, spelling and a sanity check on your arithmetic are fine. The three judgment calls are the ones being examined, and an assistant that makes them has taken the exam.

The clock breaks into seven blocks, and the shape of the split is the finding. Ten minutes to read the brief and write the audience, decision, medium and takeaway. Twenty-five to interrogate the dataset. Thirty to compute the numbers the decision turns on. Twenty to build the ghost deck on paper and make the cut. Thirty to build the exhibits. Twenty-five to build the slides and write the titles. Ten to self-review, write the delivery notes and the rationale.

Eighty-five minutes, 56.7% of the clock, run before a slide exists. Fifty-five minutes, 36.7%, run inside the tool. The last ten are the ones candidates sacrifice when they are behind, and sacrificing them is what converts a deck that would have scored 0.92 into one that scores 0.84, because the self-review pass is where four cheap defects get caught for the price of one.

Write the block times on the paper in front of you before you start, with the wall-clock time each block ends. A clock you have to compute at minute sixty is a clock you will not consult.

2. The first ten minutes: audience, decision, medium, takeaway

Ten minutes buys you four lines and a provisional sentence. It does not buy you a plan, and trying to write one here is the commonest way to lose fifteen minutes.

Line one is the audience, in three parts you already know how to write: what they know, what they are deciding, what they are afraid of. For the Tarangini board: they know the business in detail and have seen the FY26 pack; they are deciding where ₹12 crore goes; the investor director is afraid that the company keeps building capacity ahead of demand, because that is what happened last year. That last clause came out of the brief's own history section, and it is worth more than the rest of the page.

Line two is the decision, written as a question the meeting can answer with a yes or a number. "How should we think about our segment mix" is not that question. "Do we build pharma capacity, food capacity, or neither, with ₹12 crore" is.

Line three is the medium and its budget, which the brief has already fixed, so the work here is converting the slot into a page count. A 25-minute board slot with a pre-read splits into about ninety seconds of opening, eight minutes of questions, a minute to close, and 14.5 minutes of walk-through. At two minutes a page that is 7.25 pages walked, so a nine-page pre-read with two pages you expect to skip is the right size. The forty-minute American slot with no pre-read splits differently: ninety seconds of opening, fifteen minutes of questions, ninety seconds to close, twenty-two minutes of walk-through, ten pages at 2.2 minutes each.

Line four is the provisional takeaway, written before you have looked at the data properly, and marked provisional in the margin so you remember that it is a hypothesis rather than a conclusion. Writing it early is worth the risk of being wrong, because a hypothesis tells you which numbers to compute first, and computing the right numbers first is how thirty minutes becomes enough. Writing it early is also how you notice, at minute fifty, that the data has just contradicted you, which is the most valuable event in the whole exercise.

The provisional takeaway on the Indian brief was this: "Food is growing three times faster than pharma, so the capital goes to food." It is a claim, it is falsifiable, it implies an action, and it is wrong. Section 3 of the reference solution shows the exact computation that broke it.

Ten minutes ends here. You do not yet know what you will recommend, and you should not pretend to.

4. The ghost deck against a clock

Twenty minutes on paper, no software, and the output is a page count with a headline on every page and a sketch under about half of them.

The page budget comes from section 2, so it is already fixed: nine pages for the Indian board, ten for the American executive committee. Fixing the count before you write the headlines is what forces the cut to happen, because an unbounded ghost deck grows to fourteen pages and then loses five of them at minute 120, when losing them is expensive.

Write the headline for every page as a full sentence with a number in it, in the order you would say them. Do not draw anything yet. When the headlines exist, read them in order, out loud if the room allows, and ask a single question: does a reader who sees only these nine sentences reach the recommendation? If the answer is no, the fault is in the sequence, and the sequence is free to change on paper.

The Indian ghost deck's first pass had eleven pages. Two were cut and the reasons were recorded, which is a scored line in the submission.

Two cuts, one page count, twenty minutes. The revenue-growth page is the interesting cut, because it is the page most candidates keep. It is accurate, it is easy to build, it opens comfortably, and its only defect is that no decision in the room turns on it. A page that survives on comfort is the exact page the page budget exists to remove.

Sketch the exhibits only for the pages where the form is not obvious. A sketch is a rectangle with axis labels and three data marks, drawn in about forty seconds. Anything more detailed and you will defend it later out of sunk cost, which is the failure the visual branch names when it says to sketch enough to know the form and not enough to fall in love with it.

Where a page needs data you have not computed yet, write the headline as a blank with the shape of the number in it: "Food capacity earns ₹____ a slot-month, against ₹____ for pharma." A blank is honest. A placeholder number is a fabricated number that will survive to the final deck about one time in four, which is why the placeholder discipline exists.

5. Choosing the exhibit set: four that carry the argument

Ten candidate exhibits, four kept, six cut. That ratio, 40% in the Indian solution and 36.4% in the American one, is close to what an experienced analyst produces, and candidates who keep seven or eight almost always have a storyline problem rather than an exhibit problem.

An exhibit earns a page when three things are true. It carries a message that appears in a headline. That message advances the decision, in the sense that a director who accepts it is closer to a yes or a no. And no other kept exhibit carries the same message better. Anything failing the third test is a duplicate, and duplicates are the most common excess in a capstone submission because both versions took real work.

Here is the Indian candidate list, scored.

Four kept exhibits, and the sequence in which they appear is itself the argument: pharma is full and food is not, the food rate is falling, food capacity therefore loses money, so the capital goes to pharma. Each exhibit is one step, and removing any one of them breaks the chain.

Two of the cuts deserve a sentence each, because they are the ones that hurt. The waterfall is a good chart. It is well suited to a revenue bridge, it would have taken eleven minutes, and it answers "where did the growth come from", which is a question the board has already had answered in the pack. The scatter is more painful, because it is the more sophisticated chart and it does show the relationship between occupancy and contribution across all twelve cells. It loses to the two-bar chart because the two-bar chart makes the same point in about two seconds and the scatter takes about fifteen, and a pre-read page that costs thirteen extra seconds of attention for no extra content is a page that has been designed for the author.

Cut exhibits do not get deleted. They go to a named appendix with an index on the last page, because the appendix contract is what lets you cut aggressively without losing the ability to answer a question. Six cut exhibits, an index that names each one in a single line, and a director who asks about the revenue bridge gets a page number rather than an apology.

One caution about the four-exhibit target, since it is a heuristic and not a rule. The number that matters is the number of distinct claims your recommendation needs, and for most capital-allocation decisions that is three or four. A brief with two independent recommendations, or one where the audience genuinely disputes the underlying data, will need more. State the count you chose and why in the rationale, and a grader who disagrees with six will still score the reasoning.

6. Building to the grid, with one focal point a page

Thirty minutes of exhibit building and twenty-five of slide building is not enough time to make design decisions, which is why you should arrive with none left to make.

The template settles the grid, the type scale, the two colours and the placeholder positions before the clock starts. A capstone attempted on a blank default template loses about eight minutes to font and margin choices that a master would have inherited, and those eight minutes come out of the self-review. Pick your template in the week before, build one dummy page on it, and check that a chart pasted into the chart placeholder lands where you expect.

One focal point a page means one place the eye is meant to land first, and everything else on the page is subordinate to it. On the contribution chart that focal point is the food bar sitting below the zero line, so the food bar is the accent colour and the pharma bar is grey. On the options chart it is the pharma bar at 12.5%, so that bar is accented and the other two are grey. Two accents on one page is two focal points, which is the same defect as two messages.

Build the exhibits in the order the argument needs them rather than in the order they are easy, because the clock will run out and you want the shortfall to land on the last exhibit rather than the first. If the appendix chart is unfinished at minute 135, the deck still works. If the contribution chart is unfinished, there is no deck.

Source and date every exhibit as you build it, in the four-word footnote form the archetypes branch specifies. Doing it afterwards costs three minutes you will not have, and an unsourced exhibit in a capital-allocation pack is the kind of thing an investor director notices and remembers.

A note on what "building to the grid" buys under time pressure, since the mechanics branch makes the case in full. A nine-page deck on a proper master has one decision about title position, one about chart size and one about the footnote, made once. The same deck built freehand has twenty-seven such decisions, and about six of them will be made inconsistently, which is the alignment rule's most common failure and the cheapest one to avoid.

7. The title sequence and the read-only-the-titles test

Write the titles after the exhibits exist and before the bodies are finished, because a title written from a finished exhibit is honest about what the exhibit actually shows.

Every title is a full sentence with a verb. Every title carries the number that makes it checkable, unless the page has no number, in which case ask whether the page belongs. And every title claims something a reasonable person could dispute, which is what separates a title from a label.

Here is the finished Indian title stack, in order, with nothing else from the deck.

1. Put all ₹12 crore into pharma capacity and raise the food rate 1.9%.
2. Pharma capacity is 92.0% full; food capacity is 74.6% full.
3. Food volume grew 63.0% while the food rate fell 3.1% to ₹780.
4. Every food capacity slot loses ₹11.27 a month; every pharma slot earns ₹500.94.
5. ₹12 crore returns 12.5% in pharma and 1.9% in food.
6. A ₹15.11 rate rise takes existing food capacity to breakeven at today's occupancy.
7. Pharma demand above 92% occupancy is the assumption that would change this answer.
8. Three decisions, three owners, first review 15 May.
9. Appendix: twelve city-segment rows, the rate build-up, and the FY26 capex trail.

One hundred and four words, twenty-eight seconds to read at 220 words a minute, and a director who reads nothing else knows the recommendation, the evidence, the arithmetic, the assumption that would break it and who does what next. Nothing else is being tested here, and writing quality is not part of it: a stack of nine clumsy sentences that closes the argument is worth more than nine elegant labels that do not.

The test finds three failure shapes, and it finds them cheaply enough that you should run it twice, once on the ghost deck and once on the built deck.

A gap is two adjacent titles between which the reader has to supply a step. Between "food capacity is 74.6% full" and "₹12 crore returns 12.5% in pharma" there is a gap, and pages 3 and 4 fill it. Take them out and the stack still reads smoothly, which is what makes gaps hard to spot: the sequence sounds fine and the argument does not close.

A leap is a title that asserts more than its own page proves. Page 7 claims that pharma demand above 92% is the critical assumption. The page has to earn that, which it does by showing what the recommendation returns across the occupancy band the new rooms might reach, from −16.0% if they stay empty to 12.5% at today's 92.0%. Without that, the title is a leap and the honest fix is to weaken it.

A duplicate is two titles making the same claim in different words. The first pass of the American stack had "Density explains most of the operating-ratio spread" on page 2 and "Terminal size, not management, drives most of the gap" on page 4, which is one claim twice. Page 4 became the rejected-view page instead, which is a different claim and earns its place.

8. Choosing the archetype, and applying its template

The archetype is chosen from the audience and the decision, and there are only ever two or three plausible candidates. Spending more than two minutes on the choice means the audience line is underspecified.

The Tarangini board is taking a capital decision, in a scheduled slot, with a pre-read, and expects the answer on page one. That is the board-deck archetype, decisions first, with the standing pack it normally receives kept out of the way. The obvious alternative is the consulting readout, which would have opened with the objective and the approach and reached the recommendation on page four. It loses because this audience already knows the objective, and four pages of context to a board that commissioned the work reads as padding.

The Harborline executive committee is taking an operating decision, in a longer slot, with no pre-read, and will argue about the analysis in the room. That is the project readout, findings before recommendation, because the finding is counter-intuitive and a committee that is told the answer before it sees the density relationship will spend twenty minutes disbelieving the answer. The obvious alternative is the board archetype, answer first. It loses on one specific ground: the recommendation contradicts a ranking the committee has been looking at for a quarter, and a contradicted audience needs the evidence before the conclusion.

That comparison is the general rule as well. Answer first is right when the audience trusts the analysis and is short of time. Evidence first is right when the answer contradicts something the audience currently believes, and the amount of evidence you front-load is proportional to how strongly they believe it.

Applying a template means filling its slots rather than redesigning it. The shipped templates carry a layout skeleton, a placeholder storyline of action titles, chart slots and a notes structure. Replace the placeholder titles with yours, drop the exhibits into the chart slots, delete the slots you do not need, and resist the pull to move anything. A template whose grid you have adjusted is no longer a template, and the eleven minutes it takes to adjust it are eleven minutes off the analysis.

Where a template does not fit, say so in the rationale and name what you changed. A brief with two independent recommendations does not fit the board template's single decision page, and splitting that page in two is a defensible change. Changing the type scale because you prefer a different one is not.

9. The self-review pass: seven rules, in the order that finds the most

Ten minutes, three artefacts to finish, and a self-review that has to be fast enough to leave time for the rationale. Run the seven rules in the order below rather than the order they are usually listed, because this order finds the expensive defects first.

1. One message. Read each page and write its claim in the margin. Two claims means two pages or one cut. This finds the most damage per minute because a page with two messages usually loses one of them entirely in the room.
2. Title is the message. Does the title state the claim you just wrote in the margin? A mismatch here means either the title or the body is wrong, and the diagnosis takes ten seconds: whichever one the data supports is the one that stays.
3. Chart choice. Does the form match the message type? A slope chart for a two-period change, a sorted bar for a comparison, a scatter for a relationship. A wrong form found now is a five-minute rebuild; found by the audience, it is the whole page.
4. Hierarchy. Is there one focal point, and does the eye land there first? Squint at the page. If two things survive the squint, you have two.
5. Colour discipline. One accent carrying the message, everything else neutral. Count the distinct colours on each page and on the deck as a whole. More than one accent hue in a nine-page deck is almost always an accident.
6. Declutter. Gridlines, borders, legends, decimal places you do not need. Direct-label anything with four or fewer series and delete the legend.
7. Alignment. Last, because it is the most visible and the least expensive, and because a page you are about to cut does not need aligning.

Score each page from zero to one on each rule and take the mean for the page's machine score. Then commit your own rubric score before you look at any published answer, because a self-score written after seeing the model answer measures nothing. The composite is 0.6 times the machine score plus 0.4 times the rubric score, and the crown gate is 0.90.

The weighting is where the argument of the whole galaxy sits, so it is worth seeing the arithmetic. From a rubric self-score of 0.80, reaching 0.90 needs a machine score of 0.9667, which is close to perfect. From 0.70 it needs 1.0333, which is impossible. And a perfect machine score of 1.00 with a rubric of 0.75 lands exactly on 0.90, so mechanics alone can reach the crown gate only if the argument is already good. A deck cannot be cleaned into a crown.

Run the arithmetic the other way and the same point lands harder. The Indian wrong-turn deck, the one that recommends food because food grew 57.9%, scored 0.91 on the machine, which is a genuinely well-built deck. Its rubric score was 0.55, because the recommendation was wrong, and the composite was 0.766. The corrected deck scored 0.9389 and 0.90 for a composite of 0.9233. On the machine alone the two decks differ by 0.03. On the thing that matters they differ by 0.157.

10. What the machine cannot score, and the rubric that catches it

Reveal the rubric only after your self-score is committed. Its twelve rows fall into three groups, and the machine has no access to any of them.

The first group is about whether the answer is right. It asks whether the recommendation is the one the data supports, whether it allocates the resource the decision is actually about, and whether the number in the headline is the number that governs. A machine can verify that a title contains a number. It cannot verify that it is the governing number, and the distinction between revenue growth and return on capacity is exactly that distinction.

The second group is about honesty. It asks whether the assumption that would break the recommendation sits on a page rather than buried in a footnote, whether the disconfirming evidence was kept, and whether the precision in the title is matched by the precision in the data. Writing "food capacity loses ₹11.27 a slot-month" claims two decimal places of precision on a figure that moves with occupancy, so the page has to show the occupancy it assumes, and a rubric row checks that it does.

The third group is about the reader. It asks whether this audience would know what to do on Monday, whether the next steps are owned and dated, whether the appendix is indexed so a question gets a page number, and whether the opening reaches the fear the audience arrived with.

None of the twelve rows can be scored by inspecting the pixels, and all twelve can be scored by a colleague in four minutes. That asymmetry is the reason the branch grades twice and the reason the rubric is worth 40%.

One warning about self-scoring, since it is where candidates lose most. Learners systematically over-score honesty and under-score the reader. Over-scoring honesty happens because you know what you meant, so a buried assumption feels stated. Under-scoring the reader happens because you have read the deck twenty times and cannot see it cold. The repair for both is the same and takes four minutes: give the deck to somebody who has not seen it, ask them what they would do on Monday, and score the rows from their answer rather than from your intention.

11. Delivery notes as a submission artefact

The delivery notes are half a page and contain two things: the opening, written out in full, and the three questions you expect with the answer to each.

The opening runs ninety seconds, which is 195 words at 130 words a minute, and it is the only part of a presentation worth writing verbatim. It has four jobs, in this order: state the decision the meeting is taking, state your recommendation, state the one number it rests on, and say how long you will take. A room that hears those four things in the first ninety seconds spends the rest of the slot testing your argument instead of trying to work out what it is.

The three anticipated questions are chosen by asking what the audience is afraid of, which you wrote in minute three. The investor director on the Tarangini board is afraid of building ahead of demand. The three questions follow directly: what evidence is there that pharma demand exists above 92%, what happens if it does not, and why should we believe a rate rise will stick when the rate fell last year. Each gets a two-sentence answer, and each answer names the appendix page that supports it.

Write the answer to the question you cannot answer as well, because there is always one. On the Indian brief it is the first question, and the honest answer names the evidence that exists, admits what it does not establish, and states the check that would settle it before the money is committed. An answer that admits a limit and proposes a test is stronger in a board room than an answer that overclaims, and it is the one candidates are most reluctant to write.

12. The submission, and the rationale that carries it

Four artefacts: the deck, the storyboard, a one-page rationale, and the delivery notes.

The storyboard is the ghost deck as it stood at minute seventy-five, including the pages you cut, with the cut reason written next to each. Submitting a clean storyboard that matches the finished deck exactly is a warning sign rather than a strength, because it means either that nothing was cut or that the storyboard was reconstructed afterwards. The cuts are the evidence of judgment.

The rationale is one page and names every judgment call, including the wrong ones. Six things belong on it: the audience and decision as you read them, the page budget and where it came from, the exhibit set with the rejected candidates and their reasons, the assumption the recommendation rests on, the data you would have asked for, and the turn you got wrong. That last item is scored positively. A rationale that reports a provisional takeaway abandoned at minute fifty scores higher than one that presents a straight line from brief to answer, because the straight line is almost never what happened and a grader can tell.

Name the archetype and the alternative you rejected. One sentence each is enough.

State the composite you expect, from your committed machine and rubric scores, before the grader computes it. Candidates whose predicted composite is within 0.03 of the actual one are demonstrating calibration, which is a separate skill from building the deck and a good predictor of who will improve fastest afterwards.

13. The feedback loop and the habit worth keeping

A capstone score is a measurement, and a measurement you do not act on is a number you paid fourteen hours for.

Three things are worth doing with the result. Read the rubric rows you scored above the grader and ask what you could not see; those rows name your blind spot and it will be the same blind spot next time. Read the machine rules you lost points on and check whether they cluster; a candidate who loses points on declutter across seven pages has a template problem rather than seven page problems. And rebuild one page, the worst one, with no time limit, so you know what the ceiling looks like when the clock is not the constraint.

The habit the whole galaxy is trying to install is smaller than any of that, and it survives a bad week. Before the software opens, put a text file on the screen with one sentence in it that somebody could disagree with. The discipline stops there. The chart grammar, the grid, the pyramid and the seven rules all exist to make that sentence survive contact with a reader, and none of them can supply it.

Two things worth accepting about what a capstone can and cannot certify. It certifies that you can run the sequence under pressure on unfamiliar material, which is the thing most working analysts cannot do and the reason the crown exists. It does not certify taste, and it does not certify that your next deck will be good, because the next deck will have a real audience with real politics and a brief that changes on Thursday. What it gives you is a floor you can return to when a deck starts going wrong, which is worth more than it sounds, because most decks go wrong in the same four places and you will now recognise all four.

{
  "skill": "Stress-testing a finished capstone submission against the rubric rows a machine cannot score",
  "use": "After the clock has stopped and your own machine and rubric scores are committed in writing, paste the title stack, the exhibit list with rejected candidates, and the recommendation into a chat assistant and ask it to act as a hostile board member. Use its objections to find gaps you cannot see cold; do not use it to revise the submission itself, which is closed once the clock stops.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "A spreadsheet for the machine and rubric scores, committed before the reveal"],
  "prompt": "You are a non-executive director reading a nine-page pre-read. Here is the full title stack, in order: <TITLES>. Here is the recommendation and the single number it rests on: <RECOMMENDATION>. Here are the four exhibits I kept and the six I cut, with my reason for each cut: <EXHIBITS>. Do five things. 1) Reading only the titles, state the argument and say where a step is missing between two adjacent titles. 2) Name the assumption the recommendation depends on and say whether the title stack states it anywhere. 3) Ask the three questions you would ask in the room, hardest first. 4) Name one cut exhibit whose absence weakens the argument, and say why. 5) Tell me which single title claims more than a nine-page deck could prove. Do not rewrite anything.",
  "verify": "Recompute every figure the assistant repeats back to you, because a fluent restatement of your own number is not a check on it: the growth rate, the return, the breakeven and the composite each get recomputed from the raw dataset, not from the deck. Where the assistant names a missing step between two titles, confirm it against the data before inserting a page, since a step that the data cannot support is a gap you should state rather than fill. Where it claims an exhibit is needed, ask which of your four kept exhibits it would drop for it, and reject any answer that keeps five. Treat every objection as a hypothesis about your deck rather than a finding about it.",
  "diy": "You must be able to run the six-check sweep, compute the decision number and write the title stack cold, with no tool beyond a calculator, inside the clock. Those three are the gated skills and an assistant that performs any of them has taken the examination: the sweep is what tells you the cover note is 42.6% wrong, the decision number is what reverses a recommendation that revenue growth would have got backwards, and the title stack is what a board actually reads. Use the assistant after the clock stops, on a submission you have already committed, or not at all.",
  "market": "IN"
}

Common mistakes & how experts think differently

1. Opening the software at minute twenty. Tempting because the analysis feels slow and the build feels like progress. Eighty-five of the 150 minutes belong before a slide exists, and candidates who invert that ratio arrive at minute ninety with eleven exhibits and no argument. The expert's tell is that they are still writing on paper at minute seventy, and they look behind schedule right up to the moment they are not.

1. Computing the headline number instead of the decision number. Revenue growth of 57.9% is a fact about how much space got filled. The board is allocating capital to build space, so the governing number is the return on a capacity slot, and on that measure food earns minus ₹11.27 a month while pharma earns ₹500.94. Experts ask what resource the decision allocates and compute the return on that resource, which is a habit rather than a formula.

1. Averaging ratios across units of different size. The simple mean of eleven terminal operating ratios is 90.04. The revenue-weighted figure is 87.79. The 2.25-point difference describes a network that does not exist, and a title carrying the wrong one is wrong by more than most rounding errors. Experts weight by whatever the ratio's denominator is, every time, without checking whether it matters.

1. Ranking on a metric that a structural variable explains. Sorting eleven terminals by operating ratio produces a ranking whose three worst are all among the four smallest, because density explains 95.5% of the spread. Acting on that ranking would have closed Rockford, which runs 1.06 points better than its own density predicts and ranks best of the eleven once size is taken out. The expert's move is to ask what else would produce this ranking before treating it as a finding.

1. Keeping the exhibit that took the longest. Sunk cost operates on exhibits more strongly than on anything else in a deck, because an exhibit is visible work. The scatter in the Indian solution is the more sophisticated chart and it loses to a two-bar chart that makes the same point in about a seventh of the time, two seconds against fifteen. Experts decide the exhibit set from the headline list, before building, and treat build time as irrelevant to the decision.

1. Skipping the cut on paper. Cutting a page at minute seventy-five costs the forty seconds it took to write its headline. Cutting the same page at minute 130 costs the twenty minutes that built it, plus the reluctance that usually keeps it in. The expert fixes the page count before writing headlines, which converts cutting from an act of will into arithmetic.

1. Treating a clean machine score as a result. A deck can score 0.91 on seven rules and recommend the wrong thing, which is what the wrong-turn deck did. The machine checks that the form fits the message you declared; it has no view on whether the message is true or needed. Experts read a clean score as permission to start the review, and the review is the part that finds the recommendation is backwards.

1. Burying the assumption that would break the recommendation. Most capital recommendations rest on one assumption the data cannot settle, and here it is the occupancy the new pharma rooms would reach, which this dataset cannot establish. Candidates hide it because stating it feels like weakening the case. It does the opposite in a board room: the director who was going to find that assumption anyway now finds it on your page, with a named test attached, and the conversation moves to the test.

1. Submitting a storyboard that matches the deck. A clean storyboard means either that nothing was cut, which is a failure of judgment, or that the storyboard was written afterwards, which is a failure of a different kind. The cuts and their reasons are the evidence that judgment happened. Experts keep the crossed-out pages and photograph the paper.

1. Presenting a straight line from brief to answer. Real analysis contains a turn. The Indian reference solution begins with the wrong recommendation and reverses it at minute fifty; the American one begins with a ranking and abandons it at minute forty-five. A rationale that reports the turn scores higher than one that conceals it, and graders can tell the difference because concealed turns leave traces in the exhibit set.

Worked examples

Two capstone briefs, each worked end to end across two examples, then a delivery artefact and a scoring comparison. Every figure comes from one computation and the Verify block at the end of the section restates each at full precision. Both companies are synthetic composites and neither appears anywhere else in the Presentation track.

Worked example 1: The cold-chain brief, minutes 0 to 50 (India, ₹)

The brief, as handed over. Tarangini Cold Chain Private Limited operates temperature-controlled warehousing in six Indian cities. It runs two kinds of room: validated pharma rooms held at 2 to 8 degrees, and chill rooms for food. The board meets in four days and will allocate ₹12.00 crore of FY27 growth capital between pharma capacity, food capacity, or neither. Four directors attend, one of them representing a minority financial investor. You have a nine-page pre-read circulated 48 hours ahead and a 25-minute slot. The cover note says food "is now a quarter of revenue and the fastest-growing part of the business". You have 150 minutes.

Minutes 0 to 10, the four lines. The audience knows the operating detail and has seen the FY26 pack. The decision is where ₹12.00 crore goes. The fear, taken from the brief's own history paragraph, is building capacity ahead of demand, because FY26 added 1,950 food slots and 750 pharma slots and the investor director has asked twice what they earned. The medium is fixed. The 25-minute slot gives about 14.5 minutes of walk-through after ninety seconds of opening, eight minutes of questions and a minute to close, which at two minutes a page is 7.25 pages walked, so nine pages is the right pre-read.

The provisional takeaway, written at minute nine and marked provisional: food is growing three times faster than pharma, so the capital goes to food.

Minutes 10 to 35, the six-check sweep. The dataset is 24 rows, six cities by two segments by two years, five columns each: capacity slots, occupied slots, revenue per occupied slot-month, variable handling cost per occupied slot-month, and cost of capacity per capacity slot-month.

Then the cover note. Food revenue is ₹6.1776 crore of ₹35.2356 crore, which is 17.5323%, not a quarter. The claim is 7.47 percentage points high, or 42.59% high in relative terms. That correction goes in the appendix, in one line, without commentary.

Minutes 35 to 50, the computation that broke the provisional takeaway. The growth figures check out. Total revenue rose 22.4926% to ₹35.2356 crore; pharma rose 16.9185% and food rose 57.9020%; food's share of revenue moved from 13.6007% to 17.5323%, a gain of 3.9316 points. Everything in the provisional takeaway is supported so far.

Then decompose the food growth. Occupied food slots went from 4,050 to 6,600, which is 62.9630%, while the rate fell from ₹805 to ₹780, which is 3.1056%. The check: 1.629630 × 0.968944 − 1 = 57.9020%, which reproduces the revenue growth exactly. Food grew by filling rooms at a falling price.

Now the number the decision turns on. Capacity, not revenue, is what ₹12.00 crore buys, so compute the return on a capacity slot.

The pharma line is 1,240 × 0.920110 − 640 = ₹500.9366 a month, or ₹6,011.24 a year for every slot the company owns. The food line is 595 × 0.745763 − 455 = −₹11.2712 a month, or −₹135.25 a year. Food capacity, in aggregate, does not cover the cost of holding it available. It is 1.8943 percentage points short of the 76.4706% occupancy at which it would.

The provisional takeaway is now dead, and it took fifteen minutes to kill.

The wrong turn, stated plainly. The deck that would have been built from the provisional takeaway is a good deck. Its opening page reads "Food is the growth engine: 57.9% against 16.9%". Its second page shows revenue by segment across two years as a stacked column. Its third shows food's share climbing 3.93 points. Its recommendation puts ₹12.00 crore into 3,750 food slots at ₹32,000 each. Every number on it is correct, and the recommendation is wrong by 10.6203 percentage points of annual return.

Read it. Two habits produce the catch and both are cheap. The first is asking what resource the decision allocates, then computing the return on that resource rather than on the thing the pack happens to report. The second is the denominator check in the sweep: the rate is per occupied slot and the capacity cost is per capacity slot, and a segment with 74.6% occupancy loses a quarter of its revenue base to that mismatch while looking healthy on every revenue view.

The checklist to apply.

1. Name the resource the decision allocates, and compute the return per unit of it.
2. Decompose any growth rate into volume and price before quoting it.
3. Check every denominator in the dataset before combining two columns.
4. Recompute any figure the brief asserts in prose.
5. Write the provisional takeaway early, and be willing to lose it.

Worked example 2: The cold-chain brief, minutes 50 to 150 (India, ₹)

Minutes 50 to 70, the ghost deck and the cut. Eleven headlines written on paper, two cut. The revenue-growth page goes because no decision turns on it and the board has the pack. The twelve-cell utilisation page goes because the pharma spread across cities is only 7.5194 points, from Kolkata at 86.6667% to Pune at 94.1860%, and nothing in the recommendation changes anywhere in that range.

Minutes 70 to 100, the exhibit set. Ten candidates, four kept, and the four are chosen because each carries one step of the argument.

The kept dot plot puts twelve points on one axis, pharma in the accent colour and food in grey, and shows pharma clustered between 86.67% and 94.19% while food runs from 66.67% in Chennai to 81.25% in Bengaluru. The kept slope chart runs two periods and two lines: pharma ₹1,390 to ₹1,450, food ₹805 to ₹780. The kept bar chart has two bars against a zero baseline, ₹500.94 and −₹11.27, with the breakeven line drawn at zero and the food bar accented because it is the finding. The kept options chart is a sorted bar of three returns on ₹12.00 crore.

Option A returns 6.5804 times option B and 10.6203 percentage points more. At today's food occupancy of 74.5763% rather than the 85.0% target, option B returns −0.4227%, which is what the 1,950 food slots added in FY26 are actually earning on their ₹6.24 crore of capex.

The second recommendation costs no capital at all. The rate that brings existing food capacity to breakeven at today's occupancy is 455 / 0.745763 + 185 = ₹795.11, a rise of ₹15.11 or 1.9376%, which is still 1.2281% below the ₹805 the company charged in FY25. It adds ₹0.1197 crore a year on the existing book.

Minutes 100 to 125, the slides and the titles. Nine pages, built on the board template, one focal point each, and the title stack written from the finished exhibits.

One hundred and four words in the stack, twenty-eight seconds to read at 220 words a minute, and every title carries a number.

Minutes 125 to 140, the self-review. Seven rules, in order, one page at a time. Three defects surfaced. Page 6 carried two claims, the rate rise and the customer conversations needed to land it, so the second moved to page 8 as an owned action. Page 4's title said "loses money" before the review and "loses ₹11.27 a month" after, because the machine's title rule wants the number and the board wants it more. Page 9's index had six entries and no page numbers, which the appendix contract requires.

The scored deck.

The deck clears the 0.90 crown gate by 0.0233. The wrong-turn deck, built from the provisional takeaway, scored 0.91 on the machine and 0.55 on the rubric for a composite of 0.7660, short of the gate by 0.1340. The machine separated the two decks by 0.03. What separated them was a rubric row asking whether the recommendation allocates the resource the decision is about.

Read it. The recommendation is worth ₹1.6225 crore a year, of which the pharma capacity earns ₹1.5028 crore against option B's ₹0.2284 crore, a gap of 6.5804 times, but that ratio is not the reason to believe it. The reason is that the underlying arithmetic has two independent routes to the same answer: food's contribution per capacity slot is negative, and food's breakeven occupancy of 76.4706% is above its actual 74.5763%. Those two statements are the same fact seen twice, and if either had come out the other way the recommendation would have needed rebuilding.

What the number does not mean. It does not mean food is a bad business. Food revenue grew 57.9% and food capacity is 1.89 points from breakeven, which a 1.94% rate rise closes. What the number means is that new food capacity, built at today's rate and filled even to the 85% occupancy the plan assumes, would earn one rupee for every 6.58 the same money earns in pharma, and at today's occupancy it would lose money outright. The 12.5234% assumes the new pharma rooms fill to today's 92.0110%, which this dataset cannot establish. Below 57.7520% occupancy option B is the better use of the money, below 51.6129% the new capacity stops covering the cost of holding it available, and rooms that stay empty return −16.0000%. Those three figures belong on the honesty page, because an assumption that carries a recommendation is priced rather than asserted.

Worked example 3: The freight brief, minutes 0 to 45 (US, $)

The brief, as handed over. Harborline Freight Systems, Inc. is a synthetic regional less-than-truckload carrier with eleven terminals in the American Midwest. The executive committee meets in three days: chief executive, finance, operations and commercial. They are choosing between a 5.9% general rate increase effective 1 March and a $14.0m terminal-automation program, and they have been looking at a ranked list of terminal operating ratios for a quarter. Forty-minute slot, no pre-read, presented live. You have 150 minutes.

Minutes 0 to 10, the four lines. The committee knows the network intimately. The decision is which of the two programs to fund, and whether to fund both. The fear is committing $14.0m to fixed assets in a soft freight market. The slot gives twenty-two minutes of walk-through after ninety seconds of opening, fifteen minutes of questions and ninety seconds to close, which at 2.2 minutes a page is ten pages.

Minutes 10 to 35, the sweep. Eleven rows, one per terminal, nine columns. Revenue per shipment is stated as $268 network-wide and varies by less than 4% across terminals, so revenue is effectively proportional to shipment count. Total daily shipments are 9,140, which over 254 operating days is 2,321,560 shipments and $622.1781m of revenue. Every operating-ratio point is therefore worth $6.2218m of operating income.

The sweep's most valuable finding is an averaging error waiting to happen. The simple mean of the eleven operating ratios is 90.0436. The revenue-weighted figure is 87.7909. The gap is 2.2527 points, or $14.0m of operating income, and the simple mean describes a network in which the smallest terminal counts as much as the largest.

Minutes 35 to 45, the ranking, and why it broke. The committee's list ranks terminals by raw operating ratio, worst first: Rockford at 96.86, Davenport at 96.12, Peoria at 95.99. The obvious storyline is to close or fix those three.

Plot operating ratio against daily shipments and the ranking dissolves. The eleven points sit on a line: OR = 104.5607 − 0.0174714 × daily shipments, which is a fall of 1.7471 operating-ratio points for every hundred extra daily shipments. The correlation is −0.9774 and the fit explains 95.5381% of the variance in operating ratio. The three worst terminals on the raw list are all among the four smallest in the network.

Rockford ranks first on the raw list and eleventh on the residual, which makes it the best-run terminal in the network by the only measure that adjusts for what management controls. Acting on the raw ranking would have closed it, along with $25.8674m of revenue and $0.8122m of operating income, and would have left Toledo, the worst offender against its own benchmark, untouched. The two lists share exactly one name, Davenport.

The wrong turn, stated plainly. Forty-five minutes in, the storyboard read: "Three terminals drag the network; automate them for $14.0m." It is the storyline the committee expected, it matches the list they have been carrying, and it is built on a ranking that measures geography.

Read it. The move that catches this is asking what else would produce this ranking. A ranking of eleven units on a ratio invites a structural explanation before a managerial one, and the test costs four minutes: plot the ratio against the most obvious scale variable and look at the fit. When 95.5% of the spread is explained by size, the ranking is a size ranking wearing an operating name.

The checklist to apply.

1. Weight every ratio by its own denominator before quoting a network figure.
2. Before treating a ranking as a finding, regress it on the obvious structural variable.
3. Report the fit statistic, so the audience can judge how much the adjustment is doing.
4. Rank on the residual when a structural variable explains most of the raw spread.
5. Name what the residual cannot tell you, which is why a terminal underperforms.

Worked example 4: The freight brief, minutes 45 to 150 (US, $)

Minutes 45 to 80, the three options priced. Each option is scored on the same measure, network operating-ratio points, and each point is worth $6.2218m.

The rate increase announces 5.9% and realises 62% of it, which is 3.658% of price, against 1.8% of volume attrition. Revenue moves by 1.03658 × 0.982 − 1 = +1.7922%. Operating expense is 68% variable with volume, so it moves by 0.68 × (−1.8%) = −1.2240%. The operating ratio goes from 87.7909 to 87.7909 × 0.987760 / 1.017922 = 85.1896, an improvement of 2.6013 points and $17.8361m of operating income, for no capital.

The automation program spends $14.0m at the three raw-worst terminals and is claimed to cut each one's operating ratio by 3.5 points. Those three carry 15.9737% of revenue, so the network improvement is 0.159737 × 3.5 = 0.5591 points, worth $3.4785m a year. That is $25.0411m of capital per network operating-ratio point and a simple payback of 4.02 years.

The targeted fix spends $1.9m closing the residual gap at Toledo and Davenport. Weighted by revenue, (980 × 2.5212 + 610 × 2.2168) / 9,140 = 0.4183 points, worth $2.6024m a year. That is $4.5425m of capital per point and a payback of 0.73 years. The automation program costs 5.5126 times as much per point of the same measure.

The recommendation is the rate increase plus the targeted fix, which together move the operating ratio from 87.7909 to 84.7714, a gain of 3.0196 points and $20.4385m, for $1.9m of capital.

Minutes 80 to 95, the sensitivity that makes it credible. The rate increase is the larger half of the recommendation and it rests on a realisation assumption, so the deck prices the assumption rather than asserting it.

The increase stops improving the operating ratio at 9.9417% realisation, or at 10.2586% volume attrition holding realisation at 62%. Both breakpoints are far outside anything the network has seen, which is the sentence that turns an assumption into a defensible one.

Minutes 95 to 130, the deck. Ten pages on the project-readout template, findings before recommendation, because the answer contradicts a list the committee has carried for a quarter.

Page 4 is the page most candidates leave out, and it is the page this audience needs most. It shows the ranking they have been carrying, states what it actually measures, and names the terminal the ranking would have cost them. A committee that sees its own belief taken seriously and then dismantled with its own data argues for three minutes. A committee told the answer without that page argues for twenty.

The scored deck. Machine mean 0.9350, rubric mean 0.9100, composite 0.9250, clearing the crown gate by 0.0250. The weakest page on the machine is the appendix index at 0.87; the weakest on the rubric is the same page at 0.80. The wrong-turn version, the one recommending automation at the three raw-worst terminals, scored 0.94 on the machine and 0.60 on the rubric for a composite of 0.8040, short of the crown gate by 0.0960 and clear of an ordinary 0.80 gate by 0.0040. It was the better-built deck of the two and it would have cost the company $14.0m.

Read it. Two routes reach the same recommendation, which is what makes it defensible under questioning. The first is the residual: Toledo and Davenport are the only two terminals materially worse than their own density predicts, worth $1.6819m and $0.9205m of operating income respectively, or $2.6024m together. The second is the cost per point: $4.5425m against $25.0411m for the same measured improvement. Neither route depends on the other, and a committee that rejects the regression still has the cost comparison.

What the number does not mean. The regression says which terminals underperform their own benchmark. It does not say why, and the deck says so on page 8 rather than guessing. It also does not say that automation is a bad technology, only that spending $14.0m of it on the three raw-worst terminals buys 0.5591 network points because those terminals carry 15.9737% of revenue. The same $14.0m aimed at Indianapolis and Columbus would be a different proposal, and the deck names it as the one the committee should ask operations to price.

Worked example 5: The ninety-second opening and the three questions (India, ₹)

The setting. The Tarangini board slot opens in a minute. Four directors, one of them the investor representative who has asked twice what last year's food capex earned. The pre-read went out 48 hours ago. Ninety seconds of speaking at 130 words a minute is 195 words, and this is the only part worth writing verbatim.

The wrong version. "Good morning, thank you all for making time. I want to walk you through the FY26 performance and then get to the capital allocation question at the end. Starting on page 2, revenue was ₹35.24 crore, up 22.5%, which was a strong year across both segments." Two problems. It postpones the decision to the end of a 25-minute slot, and it opens with a number that supports nothing the board is deciding.

The corrected version, written out. "You are deciding today where ₹12 crore of FY27 growth capital goes. My recommendation is all of it into pharma capacity, and separately a 1.9% rise in the food rate, which needs no capital at all. The number that drives it is on page 4. Every pharma slot we own earns ₹501 a month after the cost of holding it available. Every food slot loses ₹11. Food revenue did grow 57.9% last year, and page 3 shows why that is consistent with the losing number: we filled food rooms by cutting the rate 3.1%. Pharma capacity is 92% full and food is 74.6% full, so the question in front of you is whether pharma demand exists above 92%, which is the one thing our data cannot prove. Page 7 is about that. I will take fourteen minutes and then stop for questions."

That is 142 words, which runs 65.5 seconds at 130 words a minute and leaves 24.5 seconds of the ninety, and it does four things in order: names the decision, states the recommendation, gives the one number, and sets the clock. It also puts the weakest part of the case on the table in the first ninety seconds, which is the move that buys the rest of the slot.

The three anticipated questions. Each chosen from the audience's fear, each answered in two sentences, each pointing at an appendix page.

Read it. The first question is the one with no clean answer, and the answer given names the evidence, admits its limit and proposes a test. Writing that answer in advance is what stops it being improvised in the room, where the temptation to overclaim is strongest.

The checklist to apply.

1. Write the opening verbatim; write nothing else verbatim.
2. State the decision before the recommendation, and the recommendation before the evidence.
3. Put the weakest assumption on the table inside the first ninety seconds.
4. Derive the three questions from the audience's fear, not from the deck's gaps.
5. Write the answer to the question you cannot answer, and include the test that would settle it.

Worked example 6: Two submissions, three points of machine score apart (US, $)

The setting. Two candidates submit against the Harborline brief. Both used the project-readout template, both wrote action titles, both sourced every exhibit. One recommends the rate increase plus the $1.9m targeted fix. The other recommends the $14.0m automation program at the three raw-worst terminals.

The machine scores. 0.9350 for the first and 0.9400 for the second. The automation deck is marginally the better-built artefact: its exhibits are cleaner, its ranked bar chart is a textbook sorted bar with direct labels, and it has one fewer table.

The rubric scores. 0.9100 and 0.6000. Four rows carry almost the whole difference.

The composites. 0.6 × 0.9350 + 0.4 × 0.9100 = 0.9250 and 0.6 × 0.9400 + 0.4 × 0.6000 = 0.8040. The first clears the crown gate. The second misses it by 0.0960 and would also have missed an ordinary 0.80 gate had its rubric come in three points lower.

The arithmetic that explains the weighting. From a rubric of 0.60, the machine score needed to reach 0.90 is (0.90 − 0.4 × 0.60) / 0.6 = 1.1000, which does not exist. From 0.70 it is 1.0333, which also does not exist. From 0.80 it is 0.9667, which is achievable and hard. Running it the other way, a perfect 1.00 machine score reaches the crown gate only with a rubric of 0.75 or better. Mechanics cannot buy a crown, and the ceiling is deliberate.

Read it. The two submissions differ by 0.005 on everything a machine can see and by $14.0m on the thing it cannot. That ratio is the argument for the whole two-part gate, and it is also the reason the rubric self-score must be committed before the reveal. A candidate who scores their own automation deck after seeing the reference solution will score it at 0.60 and learn nothing; a candidate who commits 0.85 beforehand and then reads the reference solution has measured their own blind spot, which is the only thing a capstone can actually teach you about yourself.

The checklist to apply.

1. Commit the machine and rubric self-scores in writing before any reveal.
2. Compute the composite yourself and predict it in the rationale.
3. Read the rows where you scored above the grader, because they name the blind spot.
4. Check whether lost machine points cluster on one rule, which is a template problem.
5. Rebuild the worst page with no clock, to see the ceiling.

Verify

Every figure printed above, at full precision, in the order it appears.

Clock: blocks of 10, 25, 30, 20, 30, 25, 10 summing to 150; before any slide 85, share 56.666667%; inside the tool 55, share 36.666667%. India slot 25 − 1.5 − 8.0 − 1.0 = 14.50000 minutes of walk-through, 14.50/2.0 = 7.25000 pages. US slot 40 − 1.5 − 15.0 − 1.5 = 22.00000 minutes, 22.00/2.2 = 10.00000 pages. Opening 90/60 × 130 = 195.00000 words.

Tarangini capacity and occupancy, FY26: pharma 18,150 and 16,700, utilisation 0.92011019 (92.0110%); food 8,850 and 6,600, utilisation 0.74576271 (74.5763%); blended 23,300/27,000 = 0.86296296 (86.2963%). FY25: pharma 14,900/17,400 = 85.63218%; food 4,050/6,900 = 58.69565%. Slots added 750 pharma, 1,950 food. Per-city pharma utilisation 91.37931, 92.10526, 91.11111, 92.64706, 86.66667, 94.18605, spread 7.51938 points; food 69.23077, 81.25000, 66.66667, 75.00000, 75.00000, 77.77778, spread 14.58333 points.

Revenue: pharma 16,700 × 1,450 × 12 = ₹29.058000 crore, food 6,600 × 780 × 12 = ₹6.177600 crore, total ₹35.235600 crore. FY25 pharma 14,900 × 1,390 × 12 = ₹24.853200 crore, food 4,050 × 805 × 12 = ₹3.912300 crore, total ₹28.765500 crore. Growth total 22.4926%, pharma 16.9185%, food 57.9020%. Food share 13.6007% to 17.5323%, change 3.9316 points. Cover-note claim 25.0000% against 17.5323%: 7.4677 points and 42.5942% relative. Rates 1,390 → 1,450 (+4.3165%) and 805 → 780 (−3.1056%). Food volume 4,050 → 6,600 (+62.9630%), pharma volume +12.0805%; check 1.629630 × 0.968944 − 1 = 57.9020%.

Unit economics: contribution per occupied slot-month 1,450 − 210 = ₹1,240.00 and 780 − 185 = ₹595.00. Per capacity slot-month 1,240 × 0.92011019 − 640 = ₹500.9366 and 595 × 0.74576271 − 455 = −₹11.2712; annualised ₹6,011.2397 and −₹135.2542. Breakeven utilisation 640/1,240 = 51.6129% and 455/595 = 76.4706%; food shortfall 1.8943 points, pharma headroom 40.3981 points. Breakeven food rate 455/0.74576271 + 185 = ₹795.1136, a rise of ₹15.1136 or 1.9376%, and 1.2281% below the FY25 rate. Segment capacity margin FY26: pharma ₹10.9104 crore (37.5470% of pharma revenue), food −₹0.1197 crore (−1.9376% of food revenue), total ₹10.7907 crore.

Capital: 12.00 crore / ₹48,000 = 2,500.00 pharma slots and /₹32,000 = 3,750.00 food slots. Option A 2,500 × 500.9366 × 12 = ₹15,028,099.17 (₹1.5028 crore, 12.5234%). Option B at 85.0% occupancy, 595 × 0.85 − 455 = ₹50.75 a slot-month, 3,750 × 50.75 × 12 = ₹2,283,750.00 (₹0.2284 crore, 1.9031%); at 74.5763% occupancy −₹507,203.39 (−0.4227%). Option C ₹9,930,359.50 (8.2753%). A less B 10.6203 points; A divided by B 6.5804. Option A against the occupancy its new rooms reach: it matches option B at (2,283,750/30,000 + 640)/1,240 = 57.7520%, covers its own cost of capacity at 51.6129%, and returns 2,500 × −640 × 12 / 12,00,00,000 = −16.0000% if they stay empty. Food price action 15.1136 × 6,600 × 12 = ₹1,197,000.00 (₹0.1197 crore). Recommendation total ₹1.6225 crore. FY26 capex 750 × 48,000 = ₹3.6000 crore and 1,950 × 32,000 = ₹6.2400 crore, total ₹9.8400 crore, 63.4146% to food; the added food slots return −0.4227%.

City revenue, FY26, ₹ crore: Pune 8.35740, Bengaluru 7.30680, Hyderabad 6.46380, Ahmedabad 5.45340, Chennai 4.40940, Kolkata 3.24480, summing to 35.23560, which ties to the segment build. Shares 23.71863, 20.73698, 18.34452, 15.47696, 12.51405, 9.20887, summing to 100.00000. Pie angles 85.38705, 74.65313, 66.04025, 55.71706, 45.05057, 33.15193 degrees; the closest pair, Hyderabad and Bengaluru, differ by 2.392467 share points and 8.612880 degrees.

Harborline: daily shipments 9,140; annual shipments 9,140 × 254 = 2,321,560; revenue 2,321,560 × 268 = $622,178,080 ($622.1781m); one operating-ratio point $6.2218m. Weighted operating ratio 87.79092; simple mean 90.04364; difference 2.25272 points. Operating income $75.9622m, margin 12.2091%. Regression OR = 104.560746 − 0.01747136 × d, or −1.747136 points per 100 daily shipments; r = −0.977436, R² = 0.955381. Residuals: Toledo +2.5212, Davenport +2.2168, Akron +0.7375, Columbus +0.1126, Peoria −0.3592, Fort Wayne −0.4207, Cedar Rapids −0.7962, Indianapolis −0.9273, Grand Rapids −0.9975, Des Moines −1.0255, Rockford −1.0616; they sum to 0.0000000000. Rockford revenue $25.8674m, operating income $0.8122m, predicted operating ratio 97.9216.

Options: realised price 5.9% × 0.62 = 3.6580%; revenue multiplier 1.03658 × 0.982 = 1.017922 (+1.7922%); expense multiplier 1 − 0.68 × 0.018 = 0.987760 (−1.2240%); pro-forma operating ratio 87.79092 × 0.987760 / 1.017922 = 85.18963, improvement 2.6013 points, operating income +$17.8361m. Sensitivity: 45% → 1.7689 pts / $11.7080m; 52% → 2.1136 / $14.2313; 62% → 2.6013 / $17.8361; 70% → 2.9874 / $20.7199; 80% → 3.4652 / $24.3247. Breakeven realisation 9.9417%; breakeven attrition 10.2586%. Automation: revenue share of the three raw-worst 15.9737%, network gain 0.159737 × 3.5 = 0.5591 points, $3.4785m, $25.0411m a point, payback 4.0247 years. Targeted fix: (980 × 2.5212 + 610 × 2.2168)/9,140 = 0.4183 points, $2.6024m, $4.5425m a point, payback 0.7301 years; ratio 5.5126. Combined 3.0196 points and $20.4385m, pro-forma operating ratio 84.7714. Revenue at risk from closing the three raw-worst $99.3851m, 15.9737% of the network. Toledo operating-income gap $1.6819m, Davenport $0.9205m, together $2.6024m.

Scores: India deck machine 0.938889, rubric 0.900000, composite 0.923333, margin 0.023333. US deck machine 0.935000, rubric 0.910000, composite 0.925000, margin 0.025000. India wrong turn 0.6 × 0.91 + 0.4 × 0.55 = 0.766000, short by 0.134000. US wrong turn 0.6 × 0.94 + 0.4 × 0.60 = 0.804000, short by 0.096000. Machine needed from a 0.55 rubric 1.133333; from 0.60, 1.100000; from 0.70, 1.033333; from 0.80, 0.966667. Rubric needed from a perfect machine 0.750000 for the crown and 0.500000 for an ordinary gate. Rubric points: 12 × 3 = 36, ordinary pass 28.8, crown pass 32.4; a rubric of 0.55 is 19.8 points and 0.60 is 21.6.

Titles and reading: India stack 104 words, 104/220 × 60 = 28.3636 seconds; US stack 128 words, 34.9091 seconds. The written opening 142 words, 142/130 × 60 = 65.5385 seconds spoken, leaving 24.4615 of the ninety.

Practice set

Work each one before reading the solution. A calculator is assumed. Numeric answers within 2% of the stated figure count as correct unless a problem says otherwise, and where a problem asks for a sentence, the test is whether a reader could disagree with it.

P1 (guided). The decision number (India). A warehouse segment charges ₹1,450 per occupied pallet slot per month, spends ₹210 per occupied slot on handling, and spends ₹640 per capacity slot per month holding the room available whether it is full or not. It owns 18,150 slots and 16,700 are occupied. A second segment charges ₹780, spends ₹185 and ₹455 on the same two lines, and has 6,600 of 8,850 slots occupied. Compute the contribution per capacity slot-month for each and state where new capital goes.

Solution. Utilisation is 16,700/18,150 = 92.0110% and 6,600/8,850 = 74.5763%. Contribution per occupied slot-month is 1,450 − 210 = ₹1,240 and 780 − 185 = ₹595. Per capacity slot-month, 1,240 × 0.920110 − 640 = ₹500.9366 and 595 × 0.745763 − 455 = −₹11.2712. New capital goes to the first segment, which earns ₹500.94 a month on every slot it owns while the second loses ₹11.27. The common wrong answer computes contribution per occupied slot and concludes both segments are profitable, which is true and irrelevant: capital buys capacity, and capacity is paid for whether it is occupied or not.

P2 (guided). The average that describes nothing (US). Eleven terminals report operating ratios of 78.30, 82.31, 84.17, 89.96, 88.94, 92.02, 91.48, 96.12, 94.33, 95.99 and 96.86, with daily shipment counts of 1,450, 1,280, 1,110, 980, 870, 760, 690, 610, 540, 470 and 380 respectively. Revenue is proportional to shipments. Compute both the simple and the revenue-weighted network operating ratio, and state the difference in operating income on $622.1781m of revenue.

Solution. The simple mean is 990.48/11 = 90.0436. The weighted figure is Σ(dᵢ × ORᵢ)/Σdᵢ = 87.7909. The difference is 2.2527 points. Each operating-ratio point is 1% of revenue, so 622.1781 × 0.010 = $6.2218m a point, and the gap is 2.2527 × 6.2218 = $14.0159m of operating income. Quoting the simple mean describes a network in which the 380-shipment terminal counts as much as the 1,450-shipment one, which is a network nobody operates.

P3 (guided). Volume against price (India). A segment's revenue grew 57.9020%. Occupied slots went from 4,050 to 6,600 and the rate per occupied slot-month went from ₹805 to ₹780. Decompose the growth, verify it reproduces the total, and write the action title.

Solution. Volume grew 6,600/4,050 − 1 = 62.9630%. Price fell 780/805 − 1 = −3.1056%. The check is 1.629630 × 0.968944 − 1 = 57.9020%, which reproduces the reported growth exactly. Title: "Food volume grew 63.0% while the food rate fell 3.1% to ₹780." A title that says only "food grew 57.9%" is true and hides the mechanism, and the mechanism is what the capital decision turns on.

P4. Ranking against residual (US). Using the fit OR = 104.5607 − 0.0174714 × d, compute the predicted operating ratio and residual for Rockford (d = 380, OR = 96.86) and for Toledo (d = 980, OR = 89.96). State which terminal a raw ranking would act on, which one a residual ranking would act on, and what that difference is worth.

Solution. Rockford's prediction is 104.5607 − 0.0174714 × 380 = 97.9216, so its residual is 96.86 − 97.9216 = −1.0616. Toledo's is 104.5607 − 0.0174714 × 980 = 87.4388, so its residual is +2.5212. A raw ranking puts Rockford first, because 96.86 is the worst number in the network, and a residual ranking puts it eleventh of eleven, which makes it the best-run terminal on the only measure that adjusts for size. Toledo ranks fourth on the raw list and first on the residual. Rockford carries 380 × 254 × 268 = $25.8674m of revenue and $0.8122m of operating income, all of which a closure decision built on the raw list would have thrown away, while Toledo's $1.6819m gap would have stayed open.

P5. The action with no capital (India). Food capacity is 8,850 slots with 6,600 occupied. Contribution per occupied slot-month is ₹595 at a rate of ₹780, and the cost of capacity is ₹455 per capacity slot-month. Find the breakeven occupancy, the rate that reaches breakeven at today's occupancy, and the annual value of that rate change on the existing book.

Solution. Breakeven occupancy is 455/595 = 76.4706%, against an actual 74.5763%, so the segment is 1.8943 points short. The rate that reaches breakeven at today's occupancy is 455/0.745763 + 185 = ₹795.1136, a rise of ₹15.1136 or 1.9376%. On 6,600 occupied slots for twelve months that is 15.1136 × 6,600 × 12 = ₹1,197,000, or ₹0.1197 crore a year, for no capital. Note the context that makes it saleable: ₹795.11 is still 1.2281% below the ₹805 charged the previous year, so it is a partial recovery rather than a new price.

P6. Cost per unit of the same measure (US). Option one spends $14.0m to cut 3.5 operating-ratio points at three terminals carrying 15.9737% of revenue. Option two spends $1.9m to close residual gaps of 2.5212 and 2.2168 points at terminals with 980 and 610 daily shipments, in a network of 9,140. Price both per network operating-ratio point on $622.1781m of revenue, and state the recommendation.

Solution. Option one delivers 0.159737 × 3.5 = 0.5591 network points, worth 0.5591 × 6.2218 = $3.4785m a year, at 14.0/0.5591 = $25.0411m a point and a payback of 4.0247 years. Option two delivers (980 × 2.5212 + 610 × 2.2168)/9,140 = 0.4183 points, worth $2.6024m, at 1.9/0.4183 = $4.5425m a point and a payback of 0.7301 years. Option one costs 5.5126 times as much per point. Fund option two, and ask operations to price the same automation aimed at the two largest terminals, where the same points would carry far more revenue.

P7. Page budget from a slot (India). A 25-minute board slot is preceded by a pre-read circulated 48 hours ahead. Allow ninety seconds to open, eight minutes for questions and a minute to close. At two minutes a page, how many pages are walked, and how many should the pre-read carry?

Solution. Walk-through is 25 − 1.5 − 8.0 − 1.0 = 14.5 minutes, which at two minutes a page is 7.25 pages. Author nine, expecting to skip two, because a pre-read carries pages the room does not need walked: the appendix index and the assumption page will be read rather than presented. Authoring exactly 7.25 pages of content and then discovering the room wants the assumption page is the failure this budget prevents.

P8. The rate-increase bridge (US). A 5.9% general rate increase realises 62% of its face value and costs 1.8% of volume. Operating expense is 68% variable with volume. Starting from an operating ratio of 87.7909, compute the pro-forma ratio, the improvement in points, and the operating income gain on $622.1781m of revenue.

Solution. Realised price is 5.9% × 0.62 = 3.658%. Revenue moves 1.03658 × 0.982 = 1.017922, so +1.7922%. Expense moves 1 − 0.68 × 0.018 = 0.987760, so −1.2240%. The new operating ratio is 87.7909 × 0.987760/1.017922 = 85.1896, an improvement of 2.6013 points. Operating income rises from $75.9622m to $93.7983m, a gain of $17.8361m, for no capital. The figure that makes it credible is the breakpoint: realisation would have to fall to 9.9417%, or attrition rise to 10.2586%, before the increase stopped improving the ratio at all.

P9. Keep four of ten (India). The candidate exhibits are: utilisation by segment and city; revenue by year and segment as a stacked column; revenue per occupied slot-month across two years as a slope chart; contribution per capacity slot-month as a two-bar chart; return on ₹12 crore under three options; revenue by city as a pie; a revenue bridge waterfall; occupancy against contribution as a scatter of twelve cells; capex by segment; headcount by city. The decision is where ₹12 crore of growth capital goes. Keep four and state what each rejected candidate would have carried.

Solution. Keep the utilisation dot plot, the rate slope chart, the contribution bar and the options bar, because they form a chain: the constraint, the price movement, the finding, the decision. The stacked column carries composition over two periods and no decision turns on it. The pie carries composition at one date across six slices whose closest pair differ by 8.6129 degrees of 360. The waterfall answers where the growth came from, which the board pack already answered. The scatter duplicates the contribution bar with more ink and a slower read. Capex by segment supports one sentence in the opening and goes to the appendix. Headcount is in the pack and touches nothing here. Six cut exhibits go to a named, indexed appendix rather than to the bin.

P10. The composite, both ways (US). Two submissions score 0.9350 and 0.9400 on the machine and 0.9100 and 0.6000 on the committed rubric. Compute both composites at a 0.6 and 0.4 blend. Then compute the machine score the second submission would have needed to reach 0.90, and say what that implies.

Solution. 0.6(0.9350) + 0.4(0.9100) = 0.9250 and 0.6(0.9400) + 0.4(0.6000) = 0.8040. The second would have needed (0.90 − 0.4 × 0.60)/0.6 = 1.1000 on the machine, which is above the maximum, so no amount of building could have rescued it. Running the same arithmetic upward, a perfect 1.00 machine score reaches 0.90 only with a rubric of 0.75 or better. The weighting is designed so that mechanics cannot buy a crown and a wrong recommendation cannot be cleaned into one.

P11. Recompute the brief (India). A brief's cover note states that a segment "is now a quarter of revenue". The segment's revenue is ₹6.1776 crore of ₹35.2356 crore. Compute the actual share, express the error two ways, and state where and how you would report it.

Solution. The actual share is 6.1776/35.2356 = 17.5323%. The cover note is 7.4677 percentage points high in absolute terms and 25.0000/17.5323 − 1 = 42.5942% high in relative terms. Report it in one line in the appendix, stating both figures and the arithmetic, without adjectives and without a slide of its own. It belongs in the appendix rather than the body because correcting the brief is not the decision the meeting is taking, and it belongs somewhere rather than nowhere because the sentence has been shaping how the board talks about the business.

P12. Repair the titles (US). Rewrite these three as titles that carry the argument, using the figures given: (a) "Terminal Performance Review", for a page showing that operating ratio falls 1.7471 points per hundred extra daily shipments with an R-squared of 0.9554; (b) "Options Analysis", for a page comparing $4.5425m and $25.0411m per operating-ratio point; (c) "Sensitivity", for a page showing that the rate increase stops helping below 9.9417% realisation. Then state the test each rewrite passes.

Solution. (a) "Operating ratio falls 1.75 points for every 100 extra daily shipments; density explains 96% of the spread." (b) "A targeted fix costs $4.5m an operating-ratio point; the automation program costs $25.0m." (c) "The rate increase still helps down to 9.9% realisation or 10.3% volume loss." Each is a full sentence with a verb, each carries the number that makes it checkable, and each could be contradicted by somebody holding different data. The rounding in (a) is honest at two significant figures because the slope is −1.747136; writing "roughly 2 points" would overstate by 14.5% and writing "1.747136" would claim a precision the eleven-point fit does not support.

Applied mini-project

The capstone itself. One hundred and fifty minutes, a brief and dataset you have not seen, solo, no assistant on the three judgment calls. Budget a further sixty minutes afterwards for the reveal, the self-scoring reconciliation and the rebuild of your worst page, which is where most of the learning happens.

Step 1: the four fixed points (10 minutes). Write the audience in three lines, the decision as a question the meeting can settle, the medium with its page budget derived from the slot, and a provisional takeaway marked provisional. Write nothing beyond those four. If you cannot fill the decision line from the brief, write the question you would ask and mark it pending; a pending line honestly marked scores better than an invented one.

Step 2: the six-check sweep (25 minutes). Produce a written page: what one row is, what ties to what, what is missing, what each column's denominator is, what is implausible, and the question the data cannot answer. Recompute at least one figure the brief asserts in prose. This page is submitted.

Step 3: the decision numbers (30 minutes). Name the resource the decision allocates and compute the return on that resource. Compute the same answer by a second, independent route wherever one exists. Where a growth rate appears, decompose it. Where a ratio is averaged across units, weight it.

Step 4: the ghost deck and the cut (20 minutes). On paper. Fix the page count first from the slot, then write a full-sentence headline with a number for every page, then read the headlines in order and check that the argument closes. Cut to the page count and write the reason next to every cut page. Photograph the paper; it is submitted as the storyboard.

Step 5: the exhibits (30 minutes). List every candidate exhibit, choose the set that carries the headline chain, and write one line for each rejected candidate naming the message it would have carried. Build in the order the argument needs, source and date each one as you build it, and send the rejects to an indexed appendix.

Step 6: build and title (25 minutes). On a template you chose before the clock started. One focal point a page. Write the titles from the finished exhibits, then read only the titles and check for gaps, leaps and duplicates.

Step 7: self-review, notes, rationale (10 minutes). Seven rules in the order given, scored per page. Commit the machine and rubric self-scores in writing before any reveal. Write the ninety-second opening verbatim, the three anticipated questions with two-sentence answers, and the one-page rationale, including the turn you got wrong and the composite you predict.

A note on the artefacts. Four files: the deck, the storyboard photograph or scan, the one-page rationale and the delivery notes. Keep the deck under whatever your mail system will carry, since most corporate systems reject attachments above roughly 20 to 25 MB and the ceiling is set per organisation (checked September 2026; confirm with your own administrator rather than assuming). Embed fonts or export to PDF, because a deck that reflows on the grader's machine loses alignment points that your own machine will not show you.

If you have no brief. Use one of your own analyses from any other galaxy, hand the raw data to somebody else, and have them write the four fixed points and hold them back until your clock starts. The unseen part is the part being examined, and self-supplied briefs almost always suit the analysis that already exists.

Rubric. Twelve rows, three points each, 36 points. Pass at 28.8 for an ordinary submission and at 32.4 for the crown, with no zero on any row. The machine score comes from the seven rules; this rubric is the 40% the machine cannot see.

Any row whose evidence is absent from the submitted artefacts scores zero regardless of how well the deck reads. A deck that recommends the wrong action cannot score above 1 on row 2, and the rows asking for the governing number and the wrong turn fall with it. A perfect machine score needs 27 of these 36 points to reach the crown gate, and the two wrong-turn decks in the worked examples came in at 19.8 and 21.6.

Reading & resources

Core (do these).

• Barbara Minto, The Pyramid Principle: the structure behind every answer-first deck, and the reference for the horizontal and vertical logic the title stack has to satisfy. Read the introduction structures before attempting a capstone, since the ninety-second opening is an SCQA in disguise. [Paid] [Intermediate]
• Gene Zelazny, Say It With Charts: the message-to-form mapping, compact enough to reread the night before. The chapter ordering is itself a good exhibit-selection checklist. [Paid] [Beginner]
• Cole Nussbaumer Knaflic, Storytelling with Data: Let's Practice: the exercise volume rather than the original, because at capstone stage what you need is reps under constraint rather than another statement of the principles. [Paid] [Beginner]
• Scott Berinato, Good Charts Workbook: short, exercise-driven, and unusually good on the moment when a chart that took two hours has to be cut. [Paid] [Beginner]
• Chip Heath and Karla Starr, Making Numbers Count: on translating a computed figure into something a room retains, which is the gap between ₹500.94 a slot-month and a director remembering it on Monday. [Paid] [Beginner]

Going deeper.

• Stephen Few, Show Me the Numbers: the reference for table and graph design, and the place to settle an argument about whether a form is defensible. [Paid] [Intermediate]
• Alberto Cairo, The Truthful Art: on what a chart is entitled to claim, which is the discipline behind the precision row in the rubric. Its treatment of uncertainty is the part most relevant here. [Paid] [Intermediate]
• Nancy Duarte, DataStory: the recommendation-first structure written for an executive audience, and a useful counterweight if your instinct is to build up to the answer. [Paid] [Beginner]
• Richard Rumelt, Good Strategy Bad Strategy: not a presentation book, and the best available treatment of why a recommendation without a diagnosis is a list of aspirations. Read the diagnosis chapters before a capital-allocation brief. [Paid] [Intermediate]
• Philip Tetlock and Dan Gardner, Superforecasting: on calibration, which is what the predicted-composite line in the rationale is training. Keeping score and adjusting is the whole mechanism. [Paid] [Intermediate]
• Jean-luc Doumont, Trees, Maps, and Theorems: dense, short and opinionated on documents, slides and speech in one frame. The chapter on effective slides is worth the book. [Paid] [Advanced]

Free and standing.

• The US Securities and Exchange Commission, A Plain English Handbook: free from sec.gov, written for disclosure documents, and the cheapest single improvement available to dense prose, including the prose in a ninety-second opening.
• W3C Web Content Accessibility Guidelines: the contrast minimum for normal-size text is 4.5 to 1, which the accent-and-grey convention in a capstone deck must still satisfy (WCAG 2.2, checked September 2026; confirm, since the guideline version advances and the ratios can move with it).

Do this, not just read. Take any deck you have already built, extract the titles alone into a plain text file, and hand that file to somebody who has not seen the deck. Ask them one question: what are you being asked to decide, and what is the recommendation? Their answer, or their inability to give one, is the read-only-the-titles test run by a real reader, and it takes them ninety seconds.

Teach it back & journal

Feynman prompt. Explain to a competent colleague, in five minutes and without using the words design, story or polish, why two decks built from the same dataset can differ by one capital decision and by only three points of machine score. Use the arithmetic. One segment's revenue grew 57.9% and its capacity earned minus ₹11.27 a slot-month, while the other grew 16.9% and earned ₹500.94, which is 6.58 times more return on the same ₹12 crore. One network's three worst terminals on a raw ratio were all among its four smallest, and the ranking that named them explained 95.5% of its own spread with size. Then make the harder point, which is that neither of those findings came from a chart. Both came from asking what resource the decision allocates and computing the return on that, before anything was drawn. If your colleague can tell you, unprompted, why the cost of capacity has a different denominator from the rate, you have taught the idea rather than the example.

Journal prompt. Take the last piece of analysis you presented to anyone, and run three passes on it with the numbers in front of you. First, name the resource the decision allocated and ask whether you computed a return on it or on something adjacent that the pack happened to report; write down which. Second, find any ratio you quoted across units of different size and recompute it weighted; write down the gap in the unit that matters, whether that is rupees, dollars or points. Third, write out your title stack alone, hand it to somebody who was not in the room, and ask them what they were being asked to decide.

Then write a paragraph on the turn. Every real analysis has a moment where the answer you expected stopped being available, and most people do not remember theirs because they did not mark it at the time. Find yours in the last piece of work, and write down what you believed at minute forty and what changed it. If you cannot find one, that is the finding, and it usually means the analysis started from a conclusion. Finish with the one habit that survives a bad week: a text file open before the software, one sentence in it that somebody could disagree with, and a rule that the software stays shut until the sentence has a number in it.

This module's flashcards and mastery quiz are wired into the app: see the node's Quiz and Reviews.