The Data Story: Narrative Arc and the Big Idea

Presentation track, branch pr-storytelling (Storytelling) · Node PR4.03 · CROWN · ~12 focused hours · Mastery gate ≥ 90%

Turant Delivery (synthetic) loses ₹10.82 on every order it fills. Across 44.17 million orders in the year to March 2026 that is ₹47.79 crore of EBITDA loss, and the arithmetic behind it is clean: ₹202.21 crore of store-level contribution against ₹250.00 crore of technology, marketing and corporate cost. The number is correct. It also describes no store the company operates.

Split the 180 dark stores by the year each one opened and the single figure dissolves into four. The 38 stores opened in FY23 clear ₹71.72 of contribution per order after paying their own rent, staff and utilities. The FY24 stores clear ₹58.34, the FY25 stores ₹33.58, and the 72 stores opened in FY26 lose ₹6.90 an order. Forty percent of the estate is carrying a loss the other sixty percent is paying for, and the blended number reports the average of a company that is profitable and a company that is not.

Both figures were in the review deck. The blended loss per order was on slide 3, in 44-point type, under the title "Unit Economics". The cohort split was on slide 27, an eleven-row table with a footnote, under the title "Store Cohort Detail". The board spent nineteen minutes on slide 3 and forty seconds on slide 27, then asked management to improve unit economics, which is an instruction nobody can act on because it names no store, no lever and no number.

The analysis was right. Every cell tied. What failed was the ordering: the finding that would have changed the decision arrived twenty-four slides after the finding that framed it, in a form that made it look like supporting detail rather than the point. That failure has nothing to do with charts or fonts, and everything to do with the sequence in which a reader meets evidence.

A data story is an ordering of true things that makes a reader arrive at a conclusion they can re-derive. That definition carries a constraint most treatments of storytelling leave out. A pitch selects the evidence that supports a conclusion the presenter already holds and puts the rest where nobody will look. A data story orders the same evidence, including the part that cuts against the conclusion, so that a sceptical reader following the order reaches the same place. Both use tension. Only one of them survives the reader who checks.

The pyramid gives an argument its structure and the title stack gives it a surface. Neither of them decides where to begin, what to hold back until the middle, or which single exhibit the room will still be quoting a week later. Those are ordering decisions, and they are what the next twelve hours are about: finding the one idea a body of analysis actually supports, building the arc that delivers it, and knowing the three points at which a story stops being true.

The standard here is higher than elsewhere in the branch. A crown project passes at a composite of 0.90 rather than 0.80, on a brief you have not seen, which means a machine score of 0.94 with a rubric self-score of 21 out of 24 clears it by 0.014 and almost nothing else does.

Learning objectives

You can:

1. Write a big idea as one sentence carrying a point of view and a stake, and reject candidates that state a fact, hedge, or describe a topic.
2. Build the three-part arc (what is, what could be, the call to act) from a dataset, with the specific number that opens each part.
3. Create tension from a gap between two true numbers rather than from adjectives, and state the honesty constraint that separates a data story from a pitch.
4. Make an abstraction land at human scale by converting an aggregate into a per-person or per-unit figure that is arithmetically identical, without inventing a fact.
5. Sequence the middle of a story as ordered contrasts, each one a pair of numbers, and defend the order against a rearrangement.
6. Separate the analytical spine from the narrative surface, and show that every beat of the surface maps to a computed figure with a named source.
7. Choose the one exhibit the audience will remember, test it against the big idea, and budget meeting minutes across the remaining exhibits.
8. Write a recommendation that is specific, owned, dated and costed, with its payback and the conditions under which it fails.
9. Place the strongest objection inside the story, answered with a number, and state what the appendix then owes the reader.
10. Apply the three tests that stop a claim from overstating (recompute, counter-sentence, population), and weaken a claim to the level the data supports.

Prerequisites & connections

Builds on. PR4.02 is the direct prerequisite. You can already write an action title, run the read-only-the-titles test on a deck, and diagnose a slide whose body proves a different claim from its title. PR4.01 supplies the pyramid, the governing thought and the SCQA opening, and all three are assumed here without being taught again. From outside the branch, PR0.01 supplies the audience profile and the medium choice, and PR2.01 supplies the chart-choice grammar that the exhibits below obey without re-deriving it. Arithmetic is the only quantitative prerequisite, though if you've worked M2.03 or M8.02 you'll recognise the cohort and mix analysis faster.

Feeds forward. PR5.01, PR5.02 and PR5.03 take this arc and pour it into fixed archetypes. The consulting readout, the board pack, the investor pitch and the research deck each impose a shape, and a shape is only useful once you can build a story to put in it. PR6.01 takes the finished story into a room, where the ordering decisions you make here determine which questions arrive and when. PR7.01 is the capstone that runs analysis to storyboard to deck on an unseen brief, and it grades the arc you build here as one of its four scored components.

Five neighbours own material that touches this one, and none of it gets rebuilt below. PR4.01 owns the pyramid: the governing thought, the key line, the MECE check on a level, and the SCQA and SCR openings. Here the subject is the sequence a reader walks through, and that's a different object from the tree the argument is stored in. PR4.02 owns titles as a sequence, the read-only-the-titles test and the repair of a body that proves the wrong claim. The storyboard below produces a title stack and hands it to that test without redefining it. PR2.03 owns the single chart's action title, small multiples and the dashboard. So the memorable exhibit comes up here only as a choice about which chart earns the room's attention, and how to draw it is left to that neighbour. PR3.02 owns the ghost deck and the dot-dash storyboard as a build technique. PR5.01 to PR5.03 own the archetypes, including where each one puts the recommendation and what its appendix is expected to contain. Any section below that touches one of those ideas names it and moves on.

Core teaching content

3. Tension without melodrama, and the honesty constraint

Tension in an analytical story is the distance between two numbers that can't both be comfortable. Language can't create it, and every attempt to create it with language weakens it.

Take two openings for the same finding. One reads "Our store economics face significant headwinds and the situation is deteriorating rapidly." The other reads "Stores opened in FY23 earn ₹71.72 an order. Stores opened last year lose ₹6.90. Forty percent of the estate is in the second group." The first has adjectives and no tension. Nothing in it can be checked, so nothing in it can surprise. The second has no adjectives and a great deal of tension, because you're now holding two numbers that need an explanation, and nobody has offered one yet.

The device is a gap, and gaps come in a small number of shapes. A gap between two groups, which is the cohort split above. A gap between two periods, which is the FY25 cohort at 402 orders a day against the FY26 cohort at 310. A gap between a claim and its evidence, which is "the new stores are just young" against a 22.9% same-age deficit. A gap between what a number was expected to be and what it is. Each shape is a pair of computed figures and a missing explanation, and the missing explanation is what pulls a reader forward.

Resolution is the explanation arriving, and it has to be the true one. Here the honesty constraint bites, and it's the line that separates this craft from persuasion.

A pitch is permitted to select. It chooses the evidence that supports a conclusion the presenter already holds, arranges it, and puts the contrary evidence somewhere a reader is unlikely to go. That isn't a moral failure in every context, and a sales presentation that did the opposite would be strange. An analytical story is judged by a different rule. Its reader is going to commit capital or people on the strength of it, and will hold you responsible for what was known at the time.

The constraint has an operational form. Every piece of evidence that materially cuts against your conclusion appears in the story, in the main line, before the recommendation, with the number attached. Not in the appendix, not in a risks slide at the end, and not as a hedge in the recommendation's own sentence. For Turant, the contrary evidence is real and it's uncomfortable. Look at the FY23 cohort's ₹88,933.60 per store per day. It's an average across survivors. Of the 49 stores opened in FY23, 11 have since closed. Including them, the cohort's average is ₹68,030.55, so the survivor figure overstates by 30.7%. That turns the maturation story from "stores mature into profit" into "78% of stores mature into profit, and the other 22% close".

That correction weakens the recommendation and improves it. It weakens the arithmetic, because the density programme now has to be measured against a cohort whose true history includes closures. And it improves the recommendation, because the plan can now name a stopping rule. Stores that fail to reach breakeven density by month twelve are closed, and the funding stops there. That's what the FY23 cohort actually did, and what the deck was implicitly pretending it hadn't.

Before you ship any story, find the sentence in it that a hostile expert would attack first, and check whether that sentence appears before or after your recommendation. If it's after, you've written a pitch. If it's before, with its number, you've written a story that survives the room.

4. The human on the other side of the number

An aggregate is unarguable and unmemorable. ₹47.79 crore of annual EBITDA loss is a fact that a board will nod at and forget by lunch, because there's no person in it and no action attached to any person.

The repair is arithmetic, and it doesn't need an anecdote. Divide the aggregate down to the level at which somebody makes a decision, and state the same fact there. Turant's 72 loss-making stores employ 16 riders each, so 1,152 riders in total. At 310 orders a day, each rider is completing 19.4 deliveries. At the breakeven density of 357.16 orders a day, that's 22.3 deliveries per rider, and the 420-order target is 26.25. Stated at that level, the entire recommendation becomes a sentence a store manager can act on. Each rider needs about three more deliveries a day to stop the store losing money, and about seven more to hit the plan.

Nothing was invented in that conversion. Turant's rider count is a fact from the operations file, and the division is a division. Only the level changed. Now the number appears where a decision is taken, and if you've ever seen a delivery rider, you can picture three more drops in a shift.

Chip Heath and Dan Heath, in Made to Stick, argue that concreteness is what makes an idea survive retelling. Their mechanism is the one at work here. An idea a person can simulate is an idea they can repeat. For a presentation, the point is narrower than the book's. You don't need a story about a rider. You need the unit at which your aggregate becomes a decision somebody actually makes.

There's a failure mode on either side of the technique, and both are common enough to name.

The first is the invented anecdote. A writer needs the number to feel human, can't find a real person, and produces a composite: "Meet Priya, a rider in Indiranagar, who completes nineteen deliveries a day and wonders whether her store will still be open next year." Priya doesn't exist, the wondering is fabricated, and a reader who discovers either will discount every figure on the slide. If you have a real rider's account with permission to use it, use it. If not, use the arithmetic. It's just as concrete, and it can't be falsified.

The second is scaling down to a level that misrepresents the mechanism. Turant's central cost of ₹250.00 crore works out at ₹56.60 per order, and it would be easy to write that each order carries ₹56.60 of head-office cost. That's arithmetically true and analytically misleading. Most of that cost doesn't vary with the next order, so a reader who takes the per-order figure as a marginal cost will reach a wrong conclusion about whether to fill it. Divide down to the unit where the decision is taken, and check that the resulting number behaves the way the decision needs it to.

5. The middle, built as a sequence of contrasts

Readers get lost in the middle of a story. The opening has a hook and the ending has a recommendation, and the part in between is where authors dump the analysis they're proud of.

Build the middle as an ordered list of contrasts. Each contrast is a pair of numbers with a claim between them, and each one moves the argument on by a single step. Choose the order so that removing any one of them breaks the chain. Turant's middle runs to five contrasts.

Read the fourth row carefully. It's the one that would be cut under time pressure, and it's doing the most work. Across the cross-sectional curve, density rises about 25 orders a day for each month of store age between month six and month eighteen. The FY26 cohort is 47.16 orders a day short of its own breakeven of 357.16. Two months of ageing closes that gap arithmetically. That sounds like an argument for patience, until you remember row 2. That cohort is starting 22.9% behind, so the curve it's on may not be the curve the previous cohort was on. Row 4 alone argues for waiting. Rows 2 and 4 together argue for measuring, then acting.

The order matters more than the content. Contrast 1 has to come first, because until the blended number is dead, every later figure is heard as a detail inside it. Contrast 2 has to come second, because the room's instinctive rebuttal to contrast 1 is "new stores always lose money", and an objection answered before it's raised doesn't become an argument. Contrast 5 has to come last because it's the decision. Offer a decision before its evidence and it gets answered from the gut.

Here's a test for the middle. Shuffle the contrasts into a different order and read the sequence again. If it still makes sense, the order wasn't carrying anything, and what you have is a list. A real chain becomes obviously wrong when it's reordered, in the same way an answer arriving before its question is wrong.

Attention limits the number of contrasts. Evidence doesn't. Five is comfortable in a twenty-minute readout. Three is right for a five-minute one. Nine is a research report, and a research report read aloud is a different failure. If you really do have nine contrasts, group them into three and demote six to the appendix, with a line in the main text saying what they'd show if asked. That's the appendix contract, and the deck archetypes formalise it.

6. The analytical spine and the narrative surface

Two things have to be true of a finished story at the same time. It has to read as a sequence a person can follow, and every beat of it has to be a computed number with a source. Those are different objects, and confusing them produces the two most common failures in data storytelling.

The narrative surface is what the reader experiences: the opening fracture, the contrasts in order, the resolution, the ask. The analytical spine is the table underneath, one row per beat, each row naming the figure, the computation and the source file. A deck with a surface and no spine is a story somebody made up. A deck with a spine and no surface is the Turant review pack you met at the start, correct in every cell and unable to change a decision.

Build the spine as an explicit artefact. Filling it in row by row is what catches the beat you can't support.

That's seven beats. Each one has its computation and its source. Add an eighth beat with no row and the story has an assertion in it, and the assertion is where a partner or a director will land.

A spine also solves the revision problem. Numbers change late: a month closes, a data error is found, a definition is corrected. When the change arrives at nine at night before a nine-in-the-morning meeting, you can repair a deck with a spine by finding the affected rows and following them to the slides they feed. A deck without one gets repaired by reading all thirty slides and hoping.

There's a further use, which is the read-back. Before you build any slides, hand the spine to the person who owns the data and ask them to check the third column. Checking the story isn't their job. They'll find the definitional error you can't see, because they know that the cohort tracker counts a store from its soft launch and the P&L counts it from its first full month, and that the two differ by about three weeks. Definitional mismatches of that kind are the most common source of a number that's wrong in a way nobody notices until the room.

7. The one exhibit the audience will remember

A reader leaves a twenty-minute readout carrying one picture, neither none nor six. Deciding in advance which picture that'll be is a design choice with real consequences, and most decks make it by accident. In those decks, the exhibit that gets remembered is the one that was hardest to build, or the one the presenter spent longest on because a question landed there.

Choose the hero exhibit by a single test. It's the chart whose message, stated aloud, is the big idea. If your big idea is that the newest 72 stores are behind and not simply young, and your hero exhibit shows revenue by month, the exhibit is beautiful and it's the wrong one.

For Turant, the hero is a scatter with orders per day on the horizontal axis and store contribution per day on the vertical, one marked point per cohort, a zero line, and one ghost point.

The zero line does the argument's work. Four of the five points are above it and one is below, so the shape of the picture is the finding. And the ghost point does the second half. At the age the FY26 cohort is now, the previous cohort was already above water, and the vertical distance between the two points is ₹7,049.02 a store a day. Every point carries a direct label, the accent colour marks the FY26 cohort and the ghost, and everything else is grey. The title reads: "At six months, last year's stores earned ₹4,910 a day; this year's lose ₹2,139."

Compare the alternatives a reader would have been given instead. A twelve-month line of blended contribution per order shows a number that describes no store. A stacked bar of revenue by cohort shows size when the argument is about economics, and it would make the FY26 cohort look like a success, since it carries 18.4% of orders. A table of all four cohorts with nine metrics contains everything, including the finding, and communicates none of it. This scatter earns the room because its geometry is the claim.

Budget the meeting around that choice. A thirty-minute slot with six exhibits gives each one five minutes if they're treated as equals. Spend nine minutes on the hero, which is 30% of the meeting, and the other five exhibits get 4.2 minutes each. You haven't lost anything there. Four of those five exist to make the hero believable, and nobody was meant to study them.

There's a plainer version of the test for when you're too close to the material. Write down the sentence you want a director to say to a colleague in the corridor afterwards. If that sentence doesn't contain a number from your hero exhibit, one of the two is wrong. For Turant the corridor sentence is "Our newest seventy-two stores are twenty-three percent behind the last cohort at the same age, so pause the forty and fix the seventy-two." That's twenty-three words, about eleven seconds to say, and every part of it is on the scatter. Twenty-three is 22.886% rounded to the precision a spoken sentence carries, which is the direction the recompute test allows.

9. The objection lives inside the story

Every analytical recommendation has one objection that decides it. Find it, and put it in the main line of the story with its number, before the ask.

Your instinct runs the other way. Objections feel like weaknesses, and weaknesses feel like things to keep out of the argument. So the standard compromise is a risks slide near the end with four bullets and no numbers. Nobody reads that slide and it protects nothing. Worse, it teaches the room that the objections are afterthoughts. When a director raises the real one in minute six, the presenter has to improvise an answer the deck already contained.

For Turant, three objections are foreseeable, and each has a number attached.

The first is that the new stores are simply young. It's the room's default explanation, and the whole story depends on displacing it. Answer it with the same-age comparison: 310 orders a day against 402 for the previous cohort at six months, a 22.9% deficit that ageing doesn't explain. That answer goes in contrast 2, in the middle of the story, well before the end.

The second is that pausing openings cedes ground to competitors. Your answer concedes the cost and prices it. Those 40 stores would add 12,400 orders a day, 10.2% of current volume, and they'd do it while losing ₹3.12 crore a year. That's a real concession, and stating it in the main line is what makes the price credible.

The third is that density programmes don't work. Here the answer is the strongest evidence available, with its limits stated. The FY25 cohort went from 402 to 610 orders a day in twelve months, a 51.7% rise, without any programme at all. Turant's plan asks for 35.5% over twelve months in stores that start 22.9% behind. It's a lower ask against a demonstrated rate, though in a different set of cities, and that caveat has to travel with it.

Once the objections are in the main line, the appendix gets a specific job. It holds the evidence a reader would demand if they wanted to check an answer you've already given, and nothing else. For this story that means the cohort tracker with all 180 stores, the density curve by month of age with its city splits, the survivorship reconciliation, and the programme's cost build. The main line owes the appendix a pointer at the moment each objection is answered. In return, the appendix owes the main line a page that actually settles the question. An appendix of 40 pages with no pointers from the main line is storage, not evidence.

10. Endings: the ask, the next step, and the sentence they repeat

Three different things belong at the end of a story. They often get collapsed into one, and so many decks finish weakly because of it.

The ask is what you want the room to approve today, stated as a decision they can take in the meeting. "Approve the pause and the ₹5.70 crore programme" is an ask. "Consider the recommendations" isn't, because there's nothing to say yes to.

The next step is what happens on Monday regardless of the ask. It exists because meetings end without decisions more often than they end with them. Name a next step that survives a deferral and the work keeps moving. For Turant, the operations team begins the route-density diagnostic in six pilot stores this month. That costs ₹0.25 crore and produces the data the full programme would need anyway.

The sentence they repeat is the one a director says to somebody who wasn't in the room. Nobody writes it on a slide as a slogan. It's whatever remains after the deck is closed, and your only control over it is to make sure the shortest true summary of your story is also the most memorable thing in it.

Compare the two candidates. The hedged version runs 29 words and contains no number: "Our unit economics remain under pressure across the network as we continue to scale, and management is focused on driving improved throughput and operating leverage over the medium term." Nobody repeats that, because there's nothing in it to hold. Turant's version runs 23 words and takes about eleven seconds to say. It carries three checkable facts: 72 stores, 23% behind at the same age, pause 40 and fix 72.

Try this before the meeting. Say your sentence to a colleague who hasn't seen the deck, wait an hour, and ask them what you said. Whatever comes back is what will be repeated. If the number survives and the recommendation doesn't, the sentence is a finding where it should be an ask, and the verb needs changing.

11. Where a story becomes an overstatement

Three tests stand between a defensible claim and one that will be taken apart. Run all three on every number that reaches a title.

Start with the recompute test. Reproduce the figure from the source at full precision, and check that the rounding in the sentence matches the rounding in the exhibit. The same-age gap is 22.88557%, which rounds correctly to 22.9%. Writing "nearly a quarter" claims 25%, which is 2.11 points and 9.24% above the computed figure, and that phrase gets chosen because it sounds larger. Rounding in the direction of your argument is the most common overstatement in analytical writing. It's also the easiest to catch, because the check is arithmetic.

Run the counter-sentence test second. Write the strongest true sentence that argues against your claim. For Turant that sentence exists, and it hurts: "The FY23 cohort's ₹88,933.60 a day is the average of the 38 stores that survived out of 49 opened, and including the 11 that closed the true cohort average is ₹68,030.55, which is 30.7% lower." If the counter-sentence is true and you can't answer it, the claim has to be weakened until it's defensible. Here the weakened claim is stronger in practice. "78% of stores mature into profit and 22% close" converts the maturation story into a portfolio story, and it produces the closure trigger the recommendation now carries.

The third is the population test. State the denominator and name who is excluded from it. Turant's cohort figures cover open stores only. Its density curve is cross-sectional, meaning it compares different stores at different ages and doesn't follow one store through time. So it assumes the cohorts are comparable when they aren't. The FY26 cohort opened in smaller cities with lower order values, which is part of why it's behind. Naming that in the main line costs a sentence, and it heads off the objection that would otherwise dismantle the argument in the room.

Three phrases signal that a story has drifted past its evidence, and they're worth scanning for by hand. A causal verb where only a correlation was measured ("the discount drove volume" when the data shows a discount and a volume figure in the same period). A universal quantifier attached to a sample ("stores mature into profit"). And a forecast presented in the same voice as a measurement, so that "the 72 stores will reach 420 orders a day" reads with the same authority as "the 72 stores run 310 orders a day today". The repair for the third is a single word. They would reach it, under a stated condition. That's what a plan is.

12. One story, three audiences

A story built on a spine adapts without being rebuilt. Between audiences you change which beat opens, how far the arc runs, and what the ask is. You never change the spine's numbers.

Take the Turant story to three rooms.

Every figure in that table already existed. The lender's runway arithmetic uses the same EBITDA and the same capital plan. Burn is ₹47.79 crore of EBITDA loss plus ₹13.60 crore of capital spending, which is ₹61.39 crore a year against ₹212.00 crore of cash, or 3.45 years. Under the plan, capital spending stops. Instead the programme costs ₹5.70 crore, so year-one burn falls to ₹53.49 crore and runway extends to 3.96 years. To the lender, the pause the board is being asked to approve is worth 6.1 months of runway, and that's the sentence the letter opens on.

City operations leads get the smallest version, and it's the one most likely to change behaviour, because they make their decision every day. Their whole story is two numbers and a target: 19.4 deliveries per rider today, 22.3 to stop losing money, about three more drops a shift.

The substance mustn't change across the three. The board doesn't get a version where the FY26 cohort is described as on track. The lender doesn't get a version where the closures are omitted. Adjusting depth and framing to a reader's decision is the craft. Adjusting the facts to a reader's mood is the thing that ends careers, and under deadline pressure the gap between the two is smaller than it looks from here.

13. The crown standard, and where an assistant helps

The crown project is graded on an unseen brief at a composite of 0.90, blending the machine score at 0.6 and your revealed rubric self-score at 0.4. That arithmetic is unforgiving, and it's worth absorbing before you build.

A machine score of 0.94 with a rubric of 21 out of 24, which is 0.875, blends to 0.914 and passes. A machine score of 0.92 needs a rubric of 0.870. That's 20.88 out of 24, so you need 21 of 24, and 20 won't do. A perfect machine score of 1.00 with a rubric of 0.75 comes out at 0.90 on the nose and passes by nothing. And a rubric of 20 out of 24, which is 0.833, needs a machine score of 0.944 to clear the gate, higher than most first builds reach.

Read together, those four cases say the machine floor can't rescue a weak argument at this level. Alignment, colour and declutter are cheap to fix. The rubric rows aren't, because they score the big idea, the arc, the honesty of the claims and the costing of the recommendation. Build the story first and format second. Under deadline, most people work in the opposite order.

Assistants are useful for this work and dangerous in specific places, and the split is sharp.

They're good at attacking a sentence. Give an assistant your big idea and ask it to state what a reasonable opponent would say, and you'll get two objections you hadn't written down. They're good at compression, turning a 60-word beat into a 20-word title without losing the figure, provided you check the figure afterwards. And they're good at the missing-question check. Paste your title stack and ask what a sceptical reader would want to know that the stack doesn't answer.

They're dangerous at three points that matter here. An assistant will supply narrative connective tissue that asserts causation the data doesn't support, because fluent prose prefers "drove" to "coincided with". It'll invent a human anecdote on request, cheerfully, and an invented rider is the fastest way there is to destroy a deck's credibility. And it'll accept your framing, including the framing that is the error, so a story built on the blended ₹10.82 will get a polished arc around a number that describes no store.

{
  "skill": "Stress-testing a data story's arc, its objections and its claim strength before the deck is built",
  "use": "Once your big idea, your ordered contrasts and your analytical spine exist on paper, paste the spine table and the title stack into an assistant and ask it to attack them in four specific ways. Use the attack to find the missing beat and the overstated claim; never let it supply a number, a cause or a person.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "A spreadsheet or Python notebook holding the spine's computations", "A plain text file for the big idea, so no slide exists yet"],
  "prompt": "Here is my big idea: <SENTENCE>. Here is my analytical spine, one row per beat with the figure, the computation and the source: <TABLE>. Here is my title stack in order: <TITLES>. Do four things and nothing else. 1) Name the single strongest objection a hostile expert would raise, and say which beat answers it or that none does. 2) Point at every claim whose verb asserts causation when the row's computation only shows association. 3) Tell me which beat could be removed without breaking the chain, which means it is decoration. 4) State the denominator each figure implicitly assumes and flag any beat where the excluded population is not named. Do not rewrite my titles, do not add numbers, and do not invent an example.",
  "verify": "Recompute every figure the assistant repeats back, from your own source, at full precision, and check that the rounding in each title matches the exhibit it sits on. Any causation flag it raises must be settled by you against the data, not by rewording: either the computation supports the causal claim or the verb changes. For every objection it names, find the row in your spine that answers it with a number; if no row exists, the objection goes into the main line with its figure before your recommendation, or the recommendation weakens. Reject outright any human example, quotation or motivation the assistant offers, because none of it came from your data.",
  "diy": "You must be able to write the big idea, order the contrasts and fill the spine cold, from a dataset, with nothing open but a text file and a calculator. That is what the crown grades and what a room tests: a director asks where a number came from and the answer has to arrive in one sentence, from memory, with its computation. An assistant that structures your story has taken the judgement you will be asked to defend, and it cannot tell you which of two true orderings makes the decision land.",
  "market": "IN"
}

Common mistakes & how experts think differently

1. Opening on the finding you're proudest of instead of the one the room already holds. The instinct is to lead with the discovery, and it fails because a reader who hasn't yet had their existing picture broken hears the discovery as a detail. An expert opens on the number the audience walked in with, states it accurately, and then breaks it. Turant's story opens on ₹10.82 an order, which is the board's own number, and spends its second beat destroying it.

1. Building tension with adjectives. "Significant headwinds", "rapidly deteriorating" and "a concerning trend" each ask the reader to feel something the data hasn't earned yet. Tension in analysis is the distance between two computed numbers, and an expert produces it by putting ₹71.72 next to −₹6.90 and saying nothing else for a beat.

1. Treating the recommendation as the conclusion of the analysis when it's the point of it. A story that spends six of nine pages on diagnosis and one on the ask has inverted the reader's interest. It usually happens because the author is confident about the diagnosis and uneasy about the ask. That unease is the signal that the recommendation needs costing. It doesn't mean it needs shortening.

1. Putting the strongest objection in a risks slide. It reads as a disclaimer and gets skipped, and it leaves the presenter improvising in minute six on a question the deck could have answered in writing. An expert answers the decisive objection in the middle of the story, with its number, before the ask, and treats the appendix as the place a reader goes to check an answer already given.

1. Inventing the human. A composite rider with a name and a worry makes the number feel concrete. It also makes the entire deck falsifiable in the worst way, because the one fabricated element contaminates everything around it. An expert reaches for arithmetic. Divide the aggregate down to the level where a decision is taken, and let 19.4 deliveries a rider a day do the work that a fictional Priya was going to do.

1. Rounding toward the argument. 22.9% becomes "nearly a quarter", ₹5.70 crore becomes "under ₹6 crore", 5.22 months becomes "about five months". Each step is small and each one runs the same direction, which is the direction that helps. Experts round to the precision the computation supports and accept the less impressive number, because the first person to recompute is usually the person deciding.

1. Letting a cross-sectional comparison speak as if it were longitudinal. The density curve compares different cohorts at different ages, and reading it as one store's path through time assumes the cohorts are comparable. They aren't, because the FY26 stores opened in smaller cities with lower order values. An expert states the assumption in the main line. It costs one sentence and removes the objection that would otherwise end the meeting.

1. Adapting the facts instead of the framing when the audience changes. Depth, opening beat and ask all change legitimately between a board, a lender and an operations team. The closure of eleven stores and the 22.9% deficit don't. The failure is rarely a deliberate lie. It's the softer version, where the uncomfortable beat gets dropped from the shorter deck because it didn't fit, and the shorter deck is the one that gets forwarded.

1. Selecting the hero exhibit by build effort. The chart that took four hours isn't the chart that carries the idea, and the room remembers whichever one the presenter dwelt on. An expert picks the hero exhibit by testing whether its message, said aloud, is the big idea. Then they spend about 30% of the meeting on it deliberately.

1. Shipping a story with no stopping rule. A plan that describes only success asks for open-ended funding, and an experienced director hears it as a request they'll have to repeat next quarter. Naming the condition under which the plan is abandoned, here a store below 357 orders a day at month twelve, is what makes the ask finite and therefore approvable.

Worked examples

Six examples, three in India and three in the United States, built on two synthetic companies whose full datasets are given below. Every figure comes from one computation, and the Verify block at the end restates each one at full precision. Where a chart is built, the numbers behind it are printed so you can rebuild it yourself.

Worked example 1: The number that describes no store (India, ₹)

The setting. Turant Delivery (synthetic) runs 180 dark stores across Indian cities and fills quick-commerce grocery orders. The year to 31 March 2026 produced 44.17 million orders, ₹202.21 crore of store-level contribution and ₹250.00 crore of technology, marketing and corporate cost, so ₹47.79 crore of EBITDA loss. Cash is ₹212.00 crore. The board meets in nine days to approve or refuse 40 new stores at ₹34.00 lakh of capital each.

The store file carries five columns per cohort, where a cohort is the financial year in which the store opened.

The wrong version. A 34-slide quarterly review, ordered by function. Slides 1 to 3 are the summary, and slide 3 carries the blended figure of −₹10.82 an order in 44-point type under the title "Unit Economics". Slides 4 to 12 are demand: orders, average order value, category mix, city mix. Slides 13 to 20 are supply: rider utilisation, delivery times, picking productivity. Slides 21 to 26 are marketing. Slide 27 is the cohort table, eleven rows and nine columns, titled "Store Cohort Detail". Slides 28 to 34 are the expansion pipeline, ending with a request to approve 40 stores.

Nothing on any slide is wrong. It's a placement problem. The finding is on slide 27, in a form that reads as reference material, and the number that frames the whole meeting is on slide 3, in a form that reads as a conclusion.

The corrected version. Compute the cohorts first, then let the arithmetic choose the opening.

Contribution per order is the gross margin on the order less the delivery and picking cost. For the FY23 cohort that is 512 × 0.220 − 26.0 = ₹86.64. Store daily contribution is that figure times the store's daily orders, less the store's own fixed cost: 1,240 × 86.64 − 18,500 = ₹88,933.60. Divide back by orders and you get contribution per order after store fixed costs. That's the number that makes the four cohorts comparable.

The estate runs 121,020 orders a day and earns ₹5,539,982.72 of store-level contribution a day, which is ₹45.78 an order before any central cost. The FY23 cohort alone produces ₹3,379,476.80 a day, which is 61.0% of the total from 21.1% of the stores. Meanwhile the FY26 cohort produces −₹154,009.44 a day from 40.0% of the stores.

So the opening changes. You state the blended −₹10.82 first, because it's what the board holds, and follow it immediately with the two numbers that dissolve it: +₹71.72 and −₹6.90. And the cohort table moves from slide 27 to slide 2, cut from nine columns to four.

The big idea, written before anything was built. "The FY26 stores are not young, they are 22.9% behind the last cohort at the same age, so opening forty more costs ₹3.12 crore a year we do not get back." The 22.9% comes from the operations file. The FY25 cohort was running 402 orders a day at six months of age against the FY26 cohort's 310, so 1 − 310/402 = 22.886%. The ₹3.12 crore is 40 × 2,139.02 × 365 = ₹3.1230 crore.

Read it. The blended number isn't a lie, and it isn't useless either. It's the right figure for one question, which is whether the company as a whole covers its costs. It's the wrong figure for the question on the agenda, which is whether to open forty more stores. A story that opened on the cohort table without first stating the blended number would have failed in a different way, because the board would have spent the meeting reconciling the table against the finance pack when it should have been deciding.

The checklist to apply.

1. Compute the aggregate the audience already holds, in the form they hold it.
2. Cut the same data by the dimension the decision runs along, which here is store age.
3. Check whether any real unit resembles the aggregate. If none does, use the aggregate as your opening and don't present it as a finding.
4. Write the big idea as one sentence with a point of view and a stake, and name the number that would falsify it.
5. Order the deck so the dissolving figure arrives on the slide after the aggregate, not twenty-four slides later.

Worked example 2: The blended margin that did not move (US, $)

The setting. Northgate Instruments (synthetic) makes precision test instruments in the United States and sells service contracts and consumables alongside them. Revenue fell from $317.0m in FY24 to $302.0m in FY26 and operating income fell from $20.26m to $10.43m. The chief executive has told the board that the business faces a demand problem in a soft capital-spending market, and has asked for a plan to defend share.

The wrong version. A trend deck: revenue by year, gross margin by year, operating income by year, three line charts, three action titles. The gross margin chart is the problem. Blended gross margin runs 40.461%, 40.292% and 40.673% across the three years, a total move of +0.212 points, and the accurate action title for that chart is that gross margin was stable. So the deck concludes that the margin is holding and the issue is volume. That's what the chief executive already believed, and all the deck does is confirm it.

The corrected version. Decompose. Gross profit fell from $128.26m to $122.83m, a fall of $5.4275m, and the fall splits cleanly into a rate effect and a volume and mix effect.

The rate effect holds FY26 revenue and applies the change in each segment's own margin: 168.0 × (0.245 − 0.310) = −$10.920m for instruments, 78.5 × (0.595 − 0.580) = +$1.1775m for service, and 55.5 × (0.630 − 0.620) = +$0.555m for consumables, so −$9.1875m in total. The volume and mix effect applies the FY24 margin to the change in each segment's revenue: (168.0 − 212.0) × 0.310 = −$13.640m, (78.5 − 64.0) × 0.580 = +$8.410m and (55.5 − 41.0) × 0.620 = +$8.990m, so +$3.760m in total. The two sum to −$5.4275m, which is the actual change, and the tie is the check that the decomposition is complete.

The blended margin was flat because two things cancelled. Instruments lost 6.5 points of margin. Service and consumables, which carry margins around 60%, grew from 33.1% of revenue to 44.4% of revenue, and the mix shift added back nearly all of what the rate loss took away. A single blended line hid both.

The second decomposition, and the finding. Instruments sold 40,000 units at an average realised price of $5,300 in FY24 and 42,000 units at $4,000 in FY26, so unit cost was 212.0 × 0.690 × 10^6 / 40,000 = $3,657 and 168.0 × 0.755 × 10^6 / 42,000 = $3,020. Three effects, holding the others constant in turn:

A cost-down programme delivered $637 a unit, worth $26.754m. Price gave away $1,300 a unit, worth $54.600m. What the price cut bought was 2,000 extra units, worth $3.286m at the old contribution. The chief executive called this a demand problem. It's a price-realisation problem with a demand story told over it, and the story survived two years because the blended margin never moved.

The action title for the hero exhibit. "A $26.8m cost reduction was handed to customers as a $54.6m price cut, and it bought 2,000 units worth $3.3m." The exhibit is a waterfall of the three effects with the total on the right, since the message is a decomposition and a waterfall is the form that shows how parts make a whole change.

What the numbers do not say. They don't prove the price cut was unnecessary. If the alternative was losing 15,000 units to a competitor, giving away $54.6m to keep them may have been right, and nothing in this data settles that. They do say that the decision was never framed as a trade, because the blended margin reported no cost. So the finding is that a $54.6m price decision was taken without anybody pricing it.

The checklist to apply.

1. Never accept a stable blended ratio as evidence that its components are stable.
2. Decompose into rate and volume, and check that the two sum to the actual change with nothing left over.
3. Where a segment moves against the rest, decompose it again into price, volume and cost.
4. State each effect in money, not in percentages, because a decision is taken in money.
5. Ask what the losing effect bought, and put that number next to it.

Worked example 3: The storyboard, and the read-only-the-titles test (India, ₹)

The setting. The Turant board pack has nine pages, circulated 48 hours ahead. The analysis from Worked example 1 is done. Nothing has been built yet, and the storyboard is written as a title stack in a text file.

The wrong version. The first attempt, written by an analyst who had the findings and no arc.

1. FY26 Performance Overview
2. Unit Economics
3. Store Cohort Analysis
4. Density and Order Trends
5. Cost Structure
6. Expansion Pipeline
7. Competitive Landscape
8. Recommendations
9. Appendix

Read those nine titles on their own, because that's the test that decides whether a storyboard is a story. None of the nine makes a claim. Taken together, the nine are a table of contents, and a reader who reads only the titles knows the topics covered and none of the findings. You could reorder titles 1, 2, 3 and 5 without anyone noticing, and that's the signature of a list. A chain wouldn't survive it.

The corrected version. The same nine pages, with the arc imposed and every title carrying a claim.

1. We lose ₹10.82 an order, and no store in the estate resembles that number.
2. The estate is two businesses: FY23 stores earn ₹71.72 an order, FY26 stores lose ₹6.90.
3. The FY26 stores are not young, they are 22.9% behind the last cohort at the same age.
4. Sixty-one percent of our store contribution comes from 21% of our stores.
5. Waiting adds 25 orders a day per month of age, which closes a 47-order gap in two months if the curve holds.
6. It may not hold: 11 of the 49 FY23 stores closed, so 78% mature and 22% do not.
7. Forty new stores cost ₹13.60 crore of capital and add ₹3.12 crore a year of losses at the current curve.
8. Pause the forty, spend ₹5.70 crore on density in the 72, and the programme repays in 5.22 months.
9. Approve the pause and the programme; close any store below 357 orders a day at month twelve.

Now read the titles alone again. They make an argument with no gaps: a number is destroyed, replaced, located, sized, tested, bounded, priced and decided. Reordering them breaks the sequence in an obvious way, because title 3 answers the objection that title 2 provokes, and title 6 undercuts title 5 deliberately.

The three movements, mapped. Pages 1 to 4 are what is, running 44.4% of the deck. Pages 5 to 7 are what could be and what it costs, at 33.3%. Pages 8 and 9 are the call to act, at 22.2%. The analyst's first instinct was six pages of what is and one of the ask, and the correction moved two pages of diagnosis into the appendix.

The beat that was cut, and why. An early draft had a tenth page on competitive expansion rates. It was cut because no line of the argument depends on it. Whether competitors are opening fast or slowly, the recommendation is the same, since the 40 stores lose money either way. A page that doesn't change the recommendation belongs in the appendix with a pointer, and here the pointer goes in the objection on page 7.

Read it. A title stack is the cheapest artefact in a deck and the most expensive one to get wrong, because every slide built under a wrong title is work that gets thrown away. Nine titles take twenty minutes to write and about four minutes to test. Nine slides take six hours. Test first.

The checklist to apply.

1. Write the stack as sentences with verbs before opening any tool.
2. Read the titles alone, in order, and ask whether an argument follows or a topic list does.
3. Check each title carries a number that makes it checkable.
4. Check that each title answers the question the previous one provokes.
5. Shuffle the order. If the stack still reads sensibly, the sequence is carrying nothing.
6. Cut any page whose title does not change the recommendation, and put a pointer to it where the reader would ask.

Worked example 4: One turnaround told twice (US, $)

The setting. Northgate's findings from Worked example 2 are complete. Seven people will read the deck before the meeting: the chief executive, the chief financial officer, three board members, the head of sales and the head of operations. Two versions get built.

Version A, the chronology. Fourteen slides averaging 165 words, ordered by how the work happened.

Version B, the argument. Nine slides averaging 118 words, ordered by what the reader needs to decide.

The measurement. Take eight questions a reader of a turnaround deck has to be able to answer, and record the slide on which each is first answered in each version. A question answered only in the appendix is recorded as position 15, one past the last slide of Version A, which is a convention favourable to the chronology.

The mean first-answer position is 11.25 in Version A and 3.375 in Version B, a ratio of 3.33. Five of the eight questions are answered by slide 3 in Version B and none are in Version A. At 165 and 118 words a slide, a reader reaches the mean first answer after 1,856.25 words in Version A and 398.25 in Version B, which at 220 words a minute is 8.44 minutes against 1.81. Across seven readers the difference is 46.39 person-minutes before anybody has asked a question. That's most of an hour of senior attention. Version A runs 2,310 words in total and Version B runs 1,062, a 54.0% reduction.

What the measurement is, and what it is not. It measures where an answer first appears, which is a property of the artefact and can be counted without judgement. It doesn't measure comprehension. Comprehension is a property of readers, and you'd need readers to establish it. A deck that answers a question on slide 1 in a form nobody understands has scored well on this measure and badly in the room. If you reported this as a comprehension result, you'd be committing the overstatement the honesty tests are for. Call it retrieval depth.

Read it. Don't mistake the chronology for laziness. It's the order in which the work was done, and it feels rigorous to the person who did it, because every step earned the next. That's the reason it survives into so many decks, even though it's the author's order and the reader needs a different one. Don't delete the chronology. Move it, because Version A's slides 1, 2, 10 and 13 all belong in an appendix where a reader who wants the method can find it.

The checklist to apply.

1. Write the eight questions your reader must be able to answer before ordering anything.
2. Record the slide on which each is first answered, and compute the mean.
3. If the mean is past the halfway point of the deck, the deck is in the author's order.
4. Move the method, the scope and the data definitions to an appendix and leave a pointer.
5. Report retrieval depth as retrieval depth, and don't call it comprehension.

Worked example 5: The hero exhibit, and the claim that had to be weakened (India, ₹)

The setting. The Turant storyboard is fixed. Six exhibits will be built for a 30-minute board slot, and one of them has to be the picture the room leaves with.

The candidates. Four exhibits could plausibly carry the story, and three of them fail a specific test.

The hero, specified. Horizontal axis is orders a day per store, running 0 to 1,300. Vertical axis is store contribution a day in rupees, running −₹10,000 to ₹100,000, with a heavy zero line. Five points, each directly labelled, no legend.

The FY26 point and the ghost point carry the accent colour, and the other three are grey. A vertical annotation connects the two accented points and reads "₹7,049 a day apart at the same age". A footnote gives the store counts, 38, 24, 46 and 72, because a share of an estate isn't readable without them. The title states the message: "At six months, last year's stores earned ₹4,910 a day; this year's lose ₹2,139."

Why a scatter and not a bar. The message is a relationship between two quantities, density and profitability, with a threshold at zero. A sorted bar of store contribution by cohort would carry the ranking and lose the relationship, so the reader couldn't see that the threshold is crossed somewhere between 310 and 610 orders a day. A bar would also make the ghost point impossible to place, since it's a cohort already on the chart, seen at a different age.

The attention budget. Thirty minutes across six exhibits is five minutes each if they're equals. The plan gives the hero nine minutes, which is 30% of the meeting, and leaves 21 minutes across the other five, or 4.2 minutes each. That allocation is written into the speaker notes, because a presenter who hasn't decided in advance will spend the time wherever the first question lands.

The claim that had to be weakened. The story as drafted said that stores mature into profitability. Run the counter-sentence test on it and the strongest true objection is right there in the company's own records. Of the 49 stores opened in FY23, 38 are open now, so 11 closed. Those 11 were running −₹4,180 a day in their last full quarter. Including them, the FY23 cohort's true average is (38 × 88,933.60 + 11 × (−4,180)) / 49 = ₹68,030.55 a day, against the survivor-only figure of ₹88,933.60. The survivor average overstates by 30.7%, and the survival rate is 77.55%.

Three things change as a result. The claim becomes "78% of stores mature into profit and 22% close", which is weaker and true. The hero exhibit gains a footnote stating that cohort averages cover open stores only. And the recommendation gains its stopping rule. A cohort that closes 22% of its stores is a portfolio being managed, and a portfolio doesn't simply ripen, so the plan should say which stores get closed and when.

Read it. The weakened claim is the more useful one. "Stores mature into profitability" invites a board to fund every store indefinitely, because time is doing the work. "78% mature and 22% close" invites the board to ask which of the 72 will be in which group. That's the question the density programme is designed to answer by month twelve. One correction, costing a sentence, bought the recommendation its most defensible feature.

The checklist to apply.

1. List every exhibit you could build and write the one-sentence message each actually carries.
2. Pick the exhibit whose message equals the big idea, and demote the rest.
3. State the layout in words before building: axes, ranges, points, colour roles, direct labels, footnote, title.
4. Run the counter-sentence test on the claim the exhibit makes.
5. Where the counter-sentence is true, weaken the claim to what the data supports and check whether the weaker version changes the recommendation. It usually improves it.
6. Budget the meeting minutes across exhibits in the speaker notes, before the room decides for you.

Worked example 6: The costed recommendation and the objection inside the story (US, $)

The setting. Northgate's diagnosis is agreed. Instruments gave away $54.6m of price to buy $3.3m of volume, while a $26.8m cost reduction covered the tracks. The question in front of the board is what to do about FY27 pricing, and the head of sales will argue against any increase.

The wrong version. "We recommend a more disciplined pricing approach in the instruments segment, supported by revised commercial incentives, to recover margin over time." No number, no owner, no date, no cost, and no statement of what happens if volume falls. A board can't approve it and a head of sales can't argue against it, so it feels safe to write.

The corrected version, costed. Recover 4.0% of average realised price on instruments, taking the average from $4,000 to $4,160, effective with the FY27 price book on 1 February 2027. Owner: the Chief Commercial Officer. The move costs $4.2m in FY27, being $2.4m of sales-compensation redesign and $1.8m to wind down the volume-rebate programme that produced the discounting.

The board needs four lines of arithmetic. Contribution per unit today is $4,000 − $3,020 = $980. After the increase it is $4,160 − $3,020 = $1,140. At held volume of 42,000 units, gross profit rises from $41.16m to $47.88m, a gain of $6.72m, which nets to $2.52m after the $4.2m of cost. This move stops paying when volume falls far enough that the higher contribution no longer covers the units lost: 980 / 1,140 = 0.8596. So volume can fall 14.04%, which is 5,895 units, before the price recovery is worse than doing nothing. To cover its own $4.2m of cost the move needs gross profit of at least $45.36m, which requires 45.36 / 47.88 = 0.9474 of the volume. A loss of more than 5.26%, or 2,211 units, leaves the company worse off in FY27 than standing still.

So there are two numbers, and the second is the one that decides the FY27 budget. Expect the head of sales to argue for the first, because it's the generous one. A fair presentation puts both on the slide and explains the difference.

The objection, placed inside the story. The head of sales has a real argument, that a 4% increase in a competitive market will cost more than 5.26% of volume. That objection appears on slide 5 of the nine, before the ask, and it's answered with three pieces of evidence. An assertion wouldn't do. Instrument units rose only 5.0% over two years while price fell 24.5%, so the volume the discount bought was small relative to the price given away. Service and consumables follow the installed base, and the price of a new instrument isn't what drives them. They grew 27.6% and now carry 66.5% of gross profit on 44.4% of revenue, so the instrument discount isn't the only thing defending the installed base. And the plan includes a monthly volume trigger. If trailing three-month units fall below 96.0% of the FY26 run rate, the price book reverts and the sales-compensation change is held.

The appendix contract. Because the objection is answered in the main line, the appendix gets a defined job: the unit and price history by quarter, the rebate programme's actual payouts by customer tier, the competitive price points the sales team has recorded, and the compensation model with its cost build. Each of those is pointed at from the slide where the objection is answered. Anything else that was in the working file stays out.

What the numbers do not say. They don't establish that customers will accept 4%. Nothing in a historical dataset can, because the company has never tested a price increase in this segment. The elasticity implied by two years of simultaneous price and volume movement confounds price with a soft market and a product refresh. What the numbers do is bound the risk. The board is approving a move that fails only if volume falls more than 5.26%, with a trigger that stops it at 4.0%. A bounded risk presented plainly is more persuasive than a forecast presented confidently, because a director who's been sold three confident forecasts will believe the fourth less.

The checklist to apply.

1. Write the recommendation with all four properties present: specific, owned, dated, costed.
2. Compute the threshold at which the recommendation stops paying, and the separate threshold at which it stops covering its own cost.
3. Name the strongest objection, and place it before the ask with the evidence that answers it.
4. Attach a trigger that stops the plan, with the metric and the level stated.
5. Give the appendix a defined job, and point at it from the slide where the reader will ask.
6. Say plainly what the data can't establish, and don't let a forecast imply it.

Verify

Every figure printed above, at full precision, in the order the examples use them.

Start with Turant's cohort economics. Contribution per order: FY23 512 × 0.220 − 26.0 = 86.64000; FY24 486 × 0.214 − 27.5 = 76.50400; FY25 448 × 0.206 − 31.0 = 61.28800; FY26 421 × 0.198 − 38.0 = 45.35800. Store contribution a day: FY23 1,240 × 86.64 − 18,500 = 88,933.60000; FY24 980 × 76.504 − 17,800 = 57,173.92000; FY25 610 × 61.288 − 16,900 = 20,485.68000; FY26 310 × 45.358 − 16,200 = −2,139.02000. Store contribution per order: 71.72065, 58.34073, 33.58308, −6.90006. Breakeven orders a day: 18,500/86.64 = 213.52724; 17,800/76.504 = 232.66757; 16,900/61.288 = 275.74729; 16,200/45.358 = 357.15860.

The estate aggregates next. Stores 38 + 24 + 46 + 72 = 180. Orders a day 47,120 + 23,520 + 28,060 + 22,320 = 121,020. Cohort contribution a day 3,379,476.80000, 1,372,174.08000, 942,341.28000, −154,009.44000, summing to 5,539,982.72000, which is 45.77741 an order. Annualised 5,539,982.72 × 365 / 10^7 = 202.20937 crore. EBITDA 202.20937 − 250.00000 = −47.79063 crore. Orders a year 121,020 × 365 = 44,172,300. Blended EBITDA per order −47.79063 × 10^7 / 44,172,300 = −10.81914. Central cost per order 250 × 10^7 / 44,172,300 = 56.59655. Blended average order value 475.32441; gross merchandise value 2,099.61724 crore. FY26 cohort share of stores 72/180 = 40.00000%; of orders 22,320/121,020 = 18.44323%. FY23 cohort share of store contribution 3,379,476.80/5,539,982.72 = 61.00158%; share of stores 38/180 = 21.11111%.

Then the same-age comparison. FY25 cohort at six months: 402 orders a day, ₹4,910.00000 a day, implied contribution per order (4,910 + 16,900)/402 = 54.25373. Gap 1 − 310/402 = 0.22886, or 22.88557%. Store-day gap 4,910.00 − (−2,139.02) = 7,049.02000. Breakeven shortfall 357.15860 − 310 = 47.15860 orders a day, an uplift of 15.21245%. Density curve slope, months 6 to 18: (610 − 310)/12 = 25.00000 orders a day per month.

The recommendation's costs and returns follow. Capex deferred 40 × 0.34 = 13.60000 crore. Programme 72 × 0.042 = 3.02400 crore plus 72 × 62,000 × 6 / 10^7 = 2.67840 crore, total 5.70240 crore. At 420 orders a day, store contribution 420 × 45.358 − 16,200 = 2,850.36000; gain per store a day 2,850.36 − (−2,139.02) = 4,989.38000; across 72 stores 359,235.36000 a day and 13.11209 crore a year. Payback 5.70240/13.11209 = 0.43490 years, 5.21876 months. Gain per store a year 0.18211 crore; stores needed for a one-year payback 5.70240/0.18211 = 31.31254, which is 43.48963% of 72, so 32 stores, 44.44444% of 72. Uplift asked 420/310 − 1 = 35.48387%; uplift the FY25 cohort achieved 610/402 − 1 = 51.74129%. Forty more stores: 40 × 2,139.02 × 365 / 10^7 = 3.12297 crore a year of losses and 40 × 310 = 12,400 orders a day, 10.24624% of current volume. EBITDA if the 72 reach 420: −47.79063 + 13.11209 = −34.67854 crore, removing 27.43653% of the loss. Burn with 40 openings 47.79063 + 13.60000 = 61.39063 crore, runway 212/61.39063 = 3.45330 years; burn under the plan 47.79063 + 5.70240 = 53.49303 crore, runway 3.96313 years; gain 0.50984 years, 6.11805 months.

Survivorship in the FY23 cohort runs as follows. Opened in FY23 49; closed 11; survival 38/49 = 77.55102%. Survivor-only average 88,933.60000; all-opened average (38 × 88,933.60 + 11 × (−4,180))/49 = 68,030.54694; overstatement 88,933.60/68,030.54694 − 1 = 30.72598%.

At human scale, the same figures divide down. Riders per store 16, so 16 × 72 = 1,152 riders across the FY26 cohort. Deliveries a rider a day 310/16 = 19.37500 today, 357.15860/16 = 22.32241 at breakeven, 420/16 = 26.25000 at the target. Extra deliveries needed 2.94741 to breakeven and 6.87500 to the target.

Turning to Northgate's segments. Gross profit FY24 65.72000 + 37.12000 + 25.42000 = 128.26000 on revenue 317.00000, a margin of 40.46057%; FY25 52.54200 + 41.60600 + 29.95200 = 124.10000 on 308.00000, 40.29221%; FY26 41.16000 + 46.70750 + 34.96500 = 122.83250 on 302.00000, 40.67301%. Blended margin move +0.21245 points against an instrument margin move of −6.50000 points. Operating income 20.26000, 13.60000, 10.43250; operating margin 6.39117%, 4.41558%, 3.45447%. Revenue change −4.73186%; gross profit change −5.42750; operating income change −9.82750, which is −48.50691%.

The two-way bridge decomposes that fall. Rate: 168.0 × (0.245 − 0.310) = −10.92000; 78.5 × (0.595 − 0.580) = +1.17750; 55.5 × (0.630 − 0.620) = +0.55500; total −9.18750. Volume and mix: (168.0 − 212.0) × 0.310 = −13.64000; (78.5 − 64.0) × 0.580 = +8.41000; (55.5 − 41.0) × 0.620 = +8.99000; total +3.76000. Sum −5.42750, equal to the actual change to nine decimal places.

The three-way instrument bridge is built from units and prices. Units 40,000 and 42,000; realised price $5,300.00 and $4,000.00; revenue checks 212.00000 and 168.00000. Unit cost 212.0 × 0.690 × 10^6/40,000 = 3,657.00 and 168.0 × 0.755 × 10^6/42,000 = 3,020.00, a reduction of $637.00 or 17.41865%. Price reduction $1,300.00 or 24.52830%; unit volume +5.00000%. Contribution per unit 1,643.00 and 980.00, a fall of 40.35301%. Volume effect (42,000 − 40,000) × 1,643 = +3.28600; price effect 42,000 × (4,000 − 5,300) = −54.60000; cost effect 42,000 × (3,657 − 3,020) = +26.75400; sum −24.56000, equal to the actual instrument gross-profit change. Gross profit given away at the FY24 margin 168.0 × (0.310 − 0.245) = 10.92000. Instrument revenue change −20.75472%.

The price recovery arithmetic closes the example. New price 4,000 × 1.04 = 4,160.00; new contribution per unit 1,140.00. Breakeven volume ratio 980/1,140 = 0.85965, so a 14.03509% loss, 5,895 units. Gross profit at held volume 42,000 × 1,140/10^6 = 47.88000, a gain of 6.72000, net 2.52000 after 4.20000 of cost. Ratio needed to cover the cost (41.16 + 4.20)/47.88 = 0.94737, so a 5.26316% loss, 2,211 units. Service and consumables 105.00000 to 134.00000, +27.61905%, moving from 33.12303% to 44.37086% of revenue and carrying 66.49095% of FY26 gross profit against instruments' 33.50905%. Instruments are 55.62914% of FY26 revenue.

Retrieval depth was measured on eight questions. First-answer positions, Version A 14, 4, 11, 9, 11, 12, 14, 15, mean 11.25000; Version B 1, 2, 3, 3, 3, 4, 6, 5, mean 3.37500; ratio 3.33333. Answered by slide 3: 0 and 5. Words to the mean first answer 11.25 × 165 = 1,856.25000 and 3.375 × 118 = 398.25000; minutes at 220 words a minute 8.43750 and 1.81023; saving 6.62727 minutes a reader, 46.39091 person-minutes across seven. Deck totals 2,310 and 1,062 words, read in 10.50000 and 4.82727 minutes, a 54.02597% reduction.

Two smaller sets remain, the attention budget and the composite. Six exhibits in 30 minutes is 5.00000 minutes each; nine on the hero is 30.00000% of the meeting and leaves 21.00000 minutes, or 4.20000 each, for the other five. Composites: 0.6(0.94) + 0.4(0.875) = 0.91400; (0.90 − 0.6 × 0.92)/0.4 = 0.87000, which is 20.88000 of 24; 0.6(1.00) + 0.4(0.75) = 0.90000; (0.90 − 0.4 × 0.83333)/0.6 = 0.94444. Rubric conversions 20/24 = 0.83333, 21/24 = 0.87500, 22/24 = 0.91667.

Claim strength was checked last. True same-age gap 22.88557%; "nearly a quarter" claims 25.00000%, which is 2.11443 points and 9.23913% above the computed figure. Correct rounding 22.9%. The corridor sentence runs 23 words and takes 10.61538 seconds at 130 words a minute; the hedged alternative runs 29 words and contains no number.

Practice set

Work each problem before reading its 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 holding the same data could disagree with it. The two datasets are the ones given in the worked examples.

P1 (guided). Kill the aggregate (India). Turant's blended EBITDA is −₹10.82 an order. Using the cohort table, compute store contribution per order for the FY23 and FY26 cohorts, and write the big idea in one sentence.

Solution. Contribution per order is AOV × margin − delivery cost, and store contribution per order is that figure less the store's fixed cost spread over its own orders. For FY23: 512 × 0.220 − 26.0 = ₹86.64, then (1,240 × 86.64 − 18,500)/1,240 = 88,933.60/1,240 = ₹71.72. For FY26: 421 × 0.198 − 38.0 = ₹45.358, then (310 × 45.358 − 16,200)/310 = −2,139.02/310 = −₹6.90. A big idea that passes the three tests: "The FY26 stores are not young, they are 22.9% behind the last cohort at the same age, so opening forty more costs ₹3.12 crore a year we do not get back." It carries a point of view, since the head of expansion holds the opposite, and it names the stake in money. A common wrong answer is "unit economics improve with store maturity", which nobody disputes and which implies no decision.

P2 (guided). Decompose a margin that did not move (US). Northgate's blended gross margin ran 40.461%, 40.292% and 40.673% across FY24 to FY26 while instrument margin fell 6.5 points. Compute the rate and volume decomposition of the $5.4275m gross-profit fall, and state in one sentence why the blended line misled.

Solution. Rate effect uses FY26 revenue times the change in each segment's margin: instruments 168.0 × (0.245 − 0.310) = −$10.920m, service 78.5 × 0.015 = +$1.1775m, consumables 55.5 × 0.010 = +$0.555m, total −$9.1875m. Volume and mix uses the change in revenue times the FY24 margin: −44.0 × 0.310 = −$13.640m, +14.5 × 0.580 = +$8.410m, +14.5 × 0.620 = +$8.990m, total +$3.760m. The two sum to −$5.4275m, which ties to the actual change and is the check that the decomposition is complete. The sentence: the blended margin held because a 6.5-point rate loss in instruments was offset by a mix shift that took service and consumables from 33.1% to 44.4% of revenue, so a single line reported the net of two large opposing moves as stability.

P3 (guided). Put the target where a decision is taken (India). The FY26 cohort runs 310 orders a day per store with 16 riders per store. Compute the breakeven density, the shortfall, and the same three figures per rider. Then write the one-sentence instruction a store manager can act on.

Solution. Breakeven density is store fixed cost over contribution per order: 16,200/45.358 = 357.16 orders a day, so the shortfall is 357.16 − 310 = 47.16 orders a day, which is a 15.21% uplift. Per rider: 310/16 = 19.38 deliveries today, 357.16/16 = 22.32 at breakeven, and 420/16 = 26.25 at the plan target, so 2.95 more drops a shift to stop losing money and 6.88 more to hit the plan. The instruction: "Each rider needs about three more deliveries a day for this store to stop losing money, and about seven more to hit the plan." Nothing was invented, since the rider count is from the operations file and the rest is division, and the figure is now at the level where the decision is taken.

P4. Price, volume and cost, separated (US). Northgate sold 40,000 instruments at $5,300 in FY24 and 42,000 at $4,000 in FY26, at gross margins of 31.0% and 24.5%. Compute the unit costs, then split the gross-profit change into price, volume and cost effects. State what the price cut bought.

Solution. Unit costs are 212.0 × 0.690 × 10^6/40,000 = $3,657 and 168.0 × 0.755 × 10^6/42,000 = $3,020. Contribution per unit is $1,643 and $980. Volume effect (42,000 − 40,000) × 1,643 = +$3.286m; price effect 42,000 × (4,000 − 5,300) = −$54.600m; cost effect 42,000 × (3,657 − 3,020) = +$26.754m. They sum to −$24.560m, which is the actual fall in instrument gross profit. The price cut bought 2,000 extra units worth $3.286m at the old contribution, in exchange for $54.6m of realised price, while a cost programme worth $26.754m disappeared into the same gap. State it in money and leave the percentages aside, because the decision is taken in money.

P5. Apply the counter-sentence test (India). The Turant draft claims that stores mature into profitability. The FY23 cohort shows 38 open stores at ₹88,933.60 a day; 49 opened and the 11 closures were running −₹4,180 a day. Write the counter-sentence, compute the correction, and restate the claim at the strength the data supports.

Solution. The counter-sentence: "That average covers only the stores that survived." Including the closures, the cohort average is (38 × 88,933.60 + 11 × (−4,180))/49 = 3,333,496.80/49 = ₹68,030.55 a day, so the survivor figure overstates by 88,933.60/68,030.55 − 1 = 30.73%, and the survival rate is 38/49 = 77.55%. The restated claim: 78% of stores mature into profit and 22% close. As a sentence it's weaker, and as a recommendation it's stronger, because it forces the plan to name a closure trigger. Without one, the plan assumes every store ripens, and a board that has funded a cohort before will ask for the trigger.

P6. Measure retrieval depth (US). A chronology deck of 14 slides at 165 words answers eight reader questions at slides 14, 4, 11, 9, 11, 12, 14 and 15, the last being the appendix. An argument deck of 9 slides at 118 words answers them at 1, 2, 3, 3, 3, 4, 6 and 5. Compute the mean first-answer position for each, the words a reader passes before the mean answer, and the person-minutes saved across seven readers at 220 words a minute. Then state what the measurement does not establish.

Solution. Means are 90/8 = 11.25 and 27/8 = 3.375, a ratio of 3.33. Words to the mean answer are 11.25 × 165 = 1,856.25 and 3.375 × 118 = 398.25, which at 220 words a minute is 8.4375 and 1.8102 minutes, a saving of 6.6273 minutes a reader and 46.39 person-minutes across seven. This measurement doesn't establish comprehension. It records where an answer appears in an artefact, which is countable. It says nothing about whether a reader understood it, which would need readers. Reporting it as comprehension would be the overstatement the honesty tests exist to catch.

P7. Cost a recommendation (India). The Turant plan pauses 40 openings at ₹34.00 lakh each and spends ₹4.2 lakh a store plus ₹62,000 a store a month for six months across the 72 FY26 stores, targeting 420 orders a day. Compute the programme cost, the annual gain at target, the payback, and how many of the 72 stores must reach the target to repay the programme within a year.

Solution. Capex deferred is 40 × 0.34 = ₹13.60 crore. Programme cost is 72 × 0.042 = ₹3.024 crore of one-time work plus 72 × 62,000 × 6/10^7 = ₹2.6784 crore of marketing, so ₹5.7024 crore. At 420 orders a day a store earns 420 × 45.358 − 16,200 = ₹2,850.36 a day against today's −₹2,139.02, a gain of ₹4,989.38 a store a day, ₹359,235.36 a day across 72 stores and ₹13.1121 crore a year. Payback is 5.7024/13.1121 = 0.4349 years, or 5.22 months. Per store the annual gain is ₹0.18211 crore, so 5.7024/0.18211 = 31.31 stores, meaning 32 of the 72, or 44.4%, must reach the target to repay the programme inside a year. Hold contribution per order at today's ₹45.358 and say so, because delivery cost per order falls as density rises, so the ₹13.11 crore is a floor, and you shouldn't call it a forecast.

P8. Two thresholds, not one (US). Northgate proposes a 4.0% increase in average realised instrument price, from $4,000 to $4,160, at a unit cost of $3,020 and 42,000 units. Implementing it costs $4.2m in FY27. Compute the volume loss at which the price move stops paying, and the volume loss at which it stops covering its own cost. State which one belongs in the recommendation.

Solution. Contribution per unit moves from $980 to $1,140. It stops paying when 980/1,140 = 0.8596 of the volume remains, so a loss of 14.04%, which is 5,895 units. To cover the $4.2m it must produce gross profit of at least 41.16 + 4.20 = $45.36m against 42,000 × 1,140 = $47.88m at held volume, so 45.36/47.88 = 0.9474 of the volume, a loss of 5.26% or 2,211 units. Both belong on the slide and the second belongs in the recommendation, because it is the threshold that decides the FY27 budget. Presenting only the 14.04% figure would be the generous rounding of a decision, since it assumes, without saying so, that the implementation is free.

P9. One spine, three rooms (India). Turant's board, its lender and its city operations leads each need the story. Write the opening beat and the ask for each, using only figures already computed, and name the one fact that must appear in all three.

Solution. The board opens on the ₹10.82 blended loss and its dissolution into +₹71.72 and −₹6.90, and is asked to approve the pause and the ₹5.70 crore programme. The lender opens on ₹202.21 crore of store-level contribution against ₹250.00 crore of central cost, with the cash arithmetic following: burn of 47.79 + 13.60 = ₹61.39 crore a year against ₹212.00 crore of cash is 3.45 years, and the pause takes burn to 47.79 + 5.70 = ₹53.49 crore and runway to 3.96 years, a gain of 6.1 months. The lender is asked to renew on current terms with the pause written in. The city operations leads open on 310 orders a day against a breakeven of 357.16, and are asked for three more drops per rider a day by 31 March. The fact that must appear in all three is the closure history: 11 of 49 FY23 stores closed, so the maturation claim is 78%, and it isn't 100%. Depth and framing change with the reader's decision. Facts don't.

P10. Pick the hero (US). Four exhibits are available for the Northgate board: a three-year revenue line, a three-year blended gross-margin line, a waterfall of the price, volume and cost effects on instruments, and a table of segment revenue and margin for three years. Your big idea is that a $26.8m cost reduction was handed over as a $54.6m price cut. Choose the hero, state the message of each rejected candidate, and give the action title.

Solution. The waterfall is the hero, because the message is a decomposition of a change into parts, which is what a waterfall shows, and its message stated aloud is the big idea. The revenue line says revenue fell 4.7%, which is true and not the point. The blended margin line says margin was stable at about 40.5%, which is the sentence that hid the finding for two years. The segment table contains everything, including all three effects, and communicates none of them because the reader has to do the arithmetic. Title: "A $26.8m cost reduction was handed to customers as a $54.6m price cut, and it bought 2,000 units worth $3.3m." One message, one exhibit, and the number in the title is checkable against the bars.

P11. Weaken three claims (India). Rewrite each of these to the strength the Turant data supports, and name the test that catches it. (a) "Nearly a quarter of our density is missing in the new cohort." (b) "Stores mature into profitability." (c) "The density programme will take the 72 stores to 420 orders a day."

Solution. (a) The computed gap is 1 − 310/402 = 22.886%, which rounds to 22.9%. "Nearly a quarter" claims 25%, which is 2.11 points and 9.24% above the figure, and the recompute test catches it. Write "22.9% behind the last cohort at the same age." (b) The counter-sentence test catches it: 11 of 49 FY23 stores closed, so write "78% of stores mature into profit and 22% close." (c) A forecast is being stated in the voice of a measurement. Write "the programme targets 420 orders a day by 30 September 2027, and stores below 357 at month twelve close", which restores the conditional and adds the stopping rule. Each repair costs a few words and removes the sentence a hostile reader would open on.

P12. The crown arithmetic (US). The crown composite is 0.6 times the machine score plus 0.4 times the revealed rubric self-score, and passes at 0.90. Compute: (a) the composite at a machine score of 0.94 and a rubric of 21 out of 24; (b) the rubric needed at a machine score of 0.92; (c) the machine score needed at a rubric of 20 out of 24. Then state what the three answers imply about the order in which you build.

Solution. (a) 0.6(0.94) + 0.4(21/24) = 0.564 + 0.350 = 0.914, which passes. (b) (0.90 − 0.6 × 0.92)/0.4 = 0.348/0.4 = 0.870, which is 20.88 of 24, so 21 of 24, because 20 falls short. (c) (0.90 − 0.4 × 0.83333)/0.6 = 0.56667/0.6 = 0.944, which is a higher machine score than most first builds reach. The implication is that the rubric rows, which score the big idea, the arc, the honesty of the claims and the costing, cannot be bought back with formatting. Alignment and colour are cheap to fix late. An argument isn't. Build the story first.

Applied mini-project

The crown project. Budget about five hours across two sittings, with at least one night between them, because the read-only-the-titles test is worthless on a storyboard you wrote twenty minutes ago.

The deliverable is two artefacts and one page of workings. A storyboard of nine to twelve titles in a text file. A deck built from it, in the Studio, with every slide scored by the seven-rule check. And an analytical spine, one row per beat, carrying the figure, the computation and the source.

Use the brief supplied with the crown challenge, which you won't have seen. If you're working outside the challenge, use the fallback brief below. Its dataset supports only one defensible big idea, plus two plausible wrong ones.

The fallback brief. Kadamba Learning (synthetic) sells an examination-preparation subscription in India. FY26 revenue is ₹64.00 crore, up 38.0% on FY25, with 214,000 paying subscribers at an average realised price of ₹2,991 a year. Gross margin is 71.0%. Sales and marketing cost ₹34.20 crore. The cohort file shows first-year retention of 61.0% for the FY24 intake, 54.0% for the FY25 intake and 47.0% for the FY26 intake, and the blended renewal rate reported to the board is 58.0% because it is weighted by an intake mix that has shifted. Customer acquisition cost was ₹1,842 in FY25 and ₹2,417 in FY26. The board is deciding whether to approve a ₹22.00 crore marketing budget for FY27. All of that is synthetic and internally consistent. Compute anything else you need from it, and state your derivations.

Step 1: find the big idea (about 45 minutes). Compute before you write. Cut the data by the dimension the decision runs along. For the fallback brief that's intake cohort, and month is the wrong cut. Write three candidate sentences, then reject two of them in writing, naming which test each one fails. Keep the survivor and write the strongest sentence anyone could say against it.

Step 2: build the analytical spine (about 60 minutes). One row per beat: the figure, the computation at full precision, and the source. A beat with no row is an assertion and does not ship. Check that at least one row is a figure that cuts against your recommendation, because a spine with no uncomfortable row means you haven't looked.

Step 3: order the middle as contrasts (about 45 minutes). Three to five contrasts, each a pair of numbers with a claim between them, in an order where each one answers the objection the previous one provokes. Then shuffle them and read the shuffled version. Write one sentence on what breaks, and if nothing breaks, the contrasts are a list and the middle needs rebuilding.

Step 4: write the storyboard (about 45 minutes). Nine to twelve titles, every one a sentence with a verb and a number. Mark which titles carry what is, what could be, and the call to act, and check the proportions: about 40% diagnosis, about 30% alternative, about 30% call, give or take a page. Run the read-only-the-titles test the next morning, not the same evening.

Step 5: choose and specify the hero exhibit (about 30 minutes). List every exhibit you could build with its one-sentence message. Pick the one whose message equals the big idea. Specify it in words before building: axes and ranges, points or series, colour roles, direct labels, footnote, title. Write the minute budget for all exhibits into the notes.

Step 6: write the recommendation (about 45 minutes). Specific, owned, dated, costed, with a payback and a stopping rule. Compute two thresholds where two exist: the level at which the recommendation stops paying, and the level at which it stops covering its own cost. Name the strongest objection and place it in the main line before the ask, with the number that answers it.

Step 7: run the three tests and build (about 90 minutes). Recompute every figure that reaches a title and check the rounding against its exhibit. Write the counter-sentence for your central claim and weaken the claim if the counter-sentence survives. State the denominator and the excluded population for every rate. Then build the deck, score each slide, and commit your rubric self-score before revealing the rubric.

The rubric has twenty-four points, and you pass at 20 with no zero on any row.

Any row whose evidence is missing scores zero regardless of how the deck reads, and a zero on any row fails the project. The composite is 0.6 times the machine score plus 0.4 times your rubric score out of 24, and the crown gate is 0.90, so a machine score of 0.92 needs 21 of 24, and 20 isn't enough. Commit the self-score before you look at the rubric's own marking. A self-score that differs from the marked score by more than 4 points is itself a finding, and the note you write about the gap is worth more than the points.

Reading & resources

Core (do these).

• Nancy Duarte, Resonate: the source of the what-is against what-could-be arc and of the case that a central idea must combine a point of view with what is at stake. Read the chapters on the big idea and on structure; the later material on empathy is worth one pass and is aimed at keynote speaking more than at a board readout. [Paid] [Intermediate]
• Nancy Duarte, DataStory: the narrower, later book, written for people who present numbers to executives, and the closest published treatment of the recommendation-first shape used here. The chapters on the executive summary and on writing recommendations repay a second reading. [Paid] [Beginner]
• Cole Nussbaumer Knaflic, Storytelling with Data: the strongest single treatment of the exhibit side of this craft, from decluttering to choosing what a reader looks at first. Its companion volume, Storytelling with Data: Let's Practice, is the better purchase if you want repetitions more than argument. [Paid] [Beginner]
• Barbara Minto, The Pyramid Principle: the structural spine that the storytelling branch assumes. Read the first part on the pyramid and the introduction structures, which is where the situation, complication and question sequence comes from. [Paid] [Intermediate]
• Chip Heath and Dan Heath, Made to Stick: the case that concreteness is what survives retelling, and the best available discussion of why an idea a reader can simulate is an idea they repeat. Read it against the warning above about inventing the human. [Paid] [Beginner]

Deeper (for the second pass).

• Alberto Cairo, How Charts Lie: the honesty side of exhibit design, written for readers more than for makers, and that's the perspective a presenter needs when running the three tests. [Paid] [Beginner]
• Scott Berinato, Good Charts: the Harvard Business Review treatment, strongest on the difference between an exploratory chart and a declarative one, which maps closely onto the spine and surface distinction. [Paid] [Intermediate]
• Jean-luc Doumont, Trees, Maps, and Theorems: dense, short and unusually rigorous about the structure of professional communication, including the idea that a message must be reconstructible by the reader from what you gave them. [Paid] [Advanced]
• Stephen Few, Show Me the Numbers: the reference work on quantitative display, useful here mainly for the chapter on tables against graphs when you are deciding what the hero exhibit should be. [Paid] [Intermediate]
• Edward Tufte, The Visual Display of Quantitative Information: the origin of most of the vocabulary the chart-craft branch uses. Read it after the practical books, not before, or the aesthetics will crowd out the grammar. [Paid] [Advanced]

Free.

• Cole Nussbaumer Knaflic's blog and its makeover archive: several hundred worked before-and-after examples with the reasoning stated, which is the closest free equivalent of the makeover drills. [Free] [Beginner]
• The Financial Times Visual Vocabulary poster: a one-page mapping from message type to chart form, published under a permissive licence, and a useful cross-check on the grammar used in the exhibits above. [Free] [Beginner]
• Robert Kosara's writing on visualisation research: careful summaries of the experimental literature on what readers actually perceive, and a good corrective to confident claims about memorability, including claims made in presentation books. [Free] [Advanced]

Teach it back & journal

Feynman prompt. Explain to a numerate colleague who has never built a deck why two people holding the same 180-store dataset can produce presentations of opposite value, without using the words story, narrative or compelling. Use the arithmetic. The blended figure is −₹10.82 an order and no cohort in the estate is within ₹3.90 of it, because the four cohorts run at +₹71.72, +₹58.34, +₹33.58 and −₹6.90. The newest cohort is 22.9% behind the previous one at the same age, and that's a fact about matched ages. It isn't a fact about youth. And the plan costs ₹5.70 crore against ₹13.11 crore a year of gain, so it repays in 5.22 months and needs 32 of 72 stores to work. Then make the harder point, which is that none of those numbers is about presentation. Each one is about which cut of the data was put in front of a decision, and in what order. If your colleague can tell you, unprompted, why the survivorship correction strengthened the recommendation instead of weakening it, then you've taught the idea and not the arithmetic.

Journal prompt. Take the last analysis you presented, or the last deck somebody sent you, and run four checks on it. First, find the aggregate at the front and ask whether any real unit in the data resembles it; write down the number and the answer. Second, hide the slides and read only the titles in order, then write one sentence on whether an argument follows or a topic list does. Third, find the strongest objection to whatever it recommends and note whether it appears before the ask, after it, or nowhere. Fourth, pick the deck's central claim and write the counter-sentence, the strongest true statement against it, then decide, without going easy on yourself, whether the claim survives unchanged.

Write the four answers down before you interpret them, because the interpretation is where self-flattery enters. Then write a paragraph on which of the failures is yours. Some people front-load diagnosis because the diagnosis is the part they trust and the recommendation is the part they might be wrong about. Some reach for adjectives when the numbers are thin, and that's a signal about the analysis. It isn't a writing problem. Some put every objection in an appendix and are then ambushed in the room by the one they knew about. Diagnose from the artefact and leave your self-image out of it, because the artefact doesn't flatter.

Finish with one habit small enough to survive a bad week. A useful one: before any deck, write the big idea and the counter-sentence in a text file, and refuse to open the software until the counter-sentence exists. It costs about six minutes, and on the occasions when the counter-sentence is stronger than the claim, it saves the six hours you were about to spend building the wrong argument.