One Message Per Chart: the Action Title, Small Multiples, Dashboards

Presentation track, branch pr-chart-craft (Chart Craft) · Node PR2.03 · CROWN · ~12 focused hours · Mastery gate ≥ 90%

Nine lines on one chart cross each other 31 times in twelve months. That is a measured count, not an impression: nine US sector indices, monthly, 36 possible pairs, and 31 places where one line passes another. A reader asked to follow one sector through that picture is doing visual search, and visual search is not thinking. They will find the line eventually. They will have forgotten the question by the time they do.

Now put the same nine series in a three-by-three grid of small panels, all on one shared scale, and something jumps out in about two seconds. Eight of the nine have the same shape. The ninth, health care, runs upside down against the other eight. It correlates at −0.84 with the average of the rest, and it's the only sector that ends the year below where it started. Nobody added any data in that rebuild. One chart became nine charts, the crossings went to zero, and a fact that was inside the first picture all along became impossible to miss.

This is the last stop on the chart-craft branch. You've already learned to choose the right form and to declutter, so both are settled before you get here. What's left is the part that decides whether an exhibit belongs in an argument at all: whether it carries a single message, whether its title says what that message is, and whether the reader's eye goes to the place that proves it.

There's a particular failure this fixes, and it's worth naming because nobody sees it except the reader. A chart can be correctly formed, cleanly drawn, fairly scaled and accessible in both themes, and still be worthless, because it was built to display data when it should have made a claim. Its title says "Contribution margin by quarter". Its author knows what that means. Its reader, who has ninety seconds and four other papers, doesn't. And the author never finds out, because nobody in a meeting says "I could not tell what your chart was for."

There are three parts, and you learn them in order. First comes the one-message test. It's a sentence you say out loud about the chart, and if the sentence needs the word "and", you're holding two charts. Second comes the action title, which is that sentence written above the exhibit. It carries a verb and, wherever the data supports it, a number the reader can check against the picture below it. Third is the layout that makes the message easy to find: one focal point, panels instead of a bundle when the series multiply, a reading order in a dashboard, a caption that works in a document.

None of this makes a weak analysis strong, and it's worth saying so before you start. An action title on a chart whose underlying claim is wrong is a lie with better typography. What stops that is the recomputation habit the quantitative side of the programme runs on. You recompute every number in a title before the title ships, and you check every claim against the scope the data actually covers. That check catches the most common form of overstatement in the decks people actually write: a percentage change computed across a sign change. Contribution margin per order that moves from −₹2.14 to ₹12.37 has swung ₹14.50 per order across zero, and across zero the ratio has no meaning at all. Calling it a 679% improvement is wrong, and so is "tripled".

This is the branch's crown, so the bar is 0.90 against the usual 0.85. Everything before this point can be checked mechanically, and the judgements here mostly can't. A machine can see whether your title is a sentence. It can't see whether the sentence is true.

Learning objectives

You can:

1. Apply the one-message test to any chart by stating its message aloud in a single clause, identify the second message that has been smuggled in, and split the exhibit rather than annotate around the problem.
2. Write an action title that carries a subject, a verb and a checkable number, at a stated precision you can defend, within the two-line ceiling a slide imposes.
3. Diagnose a label title by its four types (filing label, dimension label, topic label, question label), state what each one costs the reader, and name the two places a label title is correct.
4. Write the title before the chart and derive the chart form from the message type, then show what changes in the exhibit when the title changes.
5. Run the three overstatement tests on a headline claim: recompute the number, check the claim's scope against the data's scope, and check the claim's strength, including the sign-change trap where a ratio has no meaning.
6. Build a single focal point into an exhibit, count the elements competing for attention before and after, and state the rule that decides which one survives.
7. Convert a crowded multi-series chart into small multiples, choose the panel order and grid shape, and defend the shared scale by computing what free scaling would exaggerate.
8. Use sparklines and in-line data in a table or a sentence, including the shared-scale rule for a sparkline column, and identify the ranking a level-only column hides.
9. Design a dashboard rather than a wall: fix the reading order, put the claim in the top-left tile, compute the reader's comparison load at a given tile count, and defend the tile budget.
10. Sequence a set of exhibits so each one advances an argument, decide what earns a page against what earns an appendix slot, and score the result against the seven-rule floor at the crown threshold of 0.90.

Prerequisites & connections

Builds on. PR2.02 directly, and its work is assumed here without being repeated: data ink, the eight things to delete first, pre-attentive attributes, the one-accent colour formula, direct labelling against legends, and the accessibility check in both themes. PR2.01 owns the form choice, so the grammar that sends a comparison across categories to a sorted bar and a change over ordered time to a line counts as settled vocabulary here. From the mechanics branch, PR1.02 supplies the grid and the type scale a well-laid-out exhibit is built on. If you haven't done the declutter work, do it first. An exhibit with a single clear message drowning in gridlines fails the crown gate on a rule taught elsewhere.

Feeds forward. PR3.02 turns exhibits into sketches inside a ghost deck, and the action titles you write here are the "dots" that storyboard uses. PR4.02 takes the single chart's title and makes it one line of a whole deck's horizontal logic, where the test is that the titles read alone tell the story with no gaps. PR4.03 puts the sequence inside a narrative arc and asks which single chart the audience will remember. The three archetype modules, PR5.01 to PR5.03, apply exhibit sequencing inside a consulting deck, a board pack and a research deck respectively, each with its own conventions about what an appendix must contain. PR7.01 is where you do all of it once, solo, under time pressure, on data nobody has seen.

Five neighbours own material that touches this one, and the boundaries are deliberate. PR2.01 owns chart-form choice in full, including the seven message types and the permanent fails. Where a form is named below, it's named and used, never re-derived. PR2.02 owns colour, decluttering and the accent formula, so the focal-point section here only asks which element gets the accent. PR4.02 owns titles as a sequence across a deck, the read-only-the-titles test and the horizontal-vertical logic pair. What's covered here is the title of one exhibit and whether its own body proves it. PR0.01 owns the audience profile, the medium choice and the one-sentence takeaway for a whole communication, which is the parent of an exhibit's message and not the same thing. M10.03 owns exhibits inside written research, where a caption and a source line do work a slide title does differently. Where a section below names one of those ideas, it names it and moves on.

Core teaching content

1. One chart, one message: the test and what fails it

The test takes four seconds and you do it out loud. Look at the exhibit and say, in one clause, what it shows. If you can't finish the sentence without "and", you've got two charts. If you can't finish it at all, you've got a table that someone drew.

Run it on a real case. Turant is a synthetic Indian quick-commerce company. Its head of finance has drawn a line chart of contribution margin per order across eight quarters: −₹38.21, −₹29.55, −₹19.48, −₹10.98, −₹8.41, −₹2.14, ₹4.57, ₹12.37. Say what it shows. "Contribution per order turned positive in the seventh quarter and reached ₹12.37." That's one clause with a verb and a number, and the picture backs it up directly. The line crosses zero between the sixth and seventh points and ends where the sentence says.

Now watch what happens when the author tries to make the same picture carry more. Five claims are true of Turant, and all five feel as if they belong on this chart:

Two of the five belong here. The other three need a different exhibit, and the reason isn't aesthetic. Claims 4 and 5 involve quantities the chart doesn't contain. Put them in the title and the body underneath can't prove it, which is the failure the machine check calls a title-body mismatch. Claim 3 is arithmetic the reader can do but hasn't been asked to do. Asking a reader to subtract two labelled points is a small tax, and it's worth charging only when the difference itself is the message. In that case, draw the difference.

The second message usually sneaks in by one of three routes, and each has a repair. It arrives as a second series with different units, the dual-axis temptation, and you repair it by indexing both to 100 or splitting the exhibit. It arrives as an annotation making a claim the picture doesn't support, and you delete the annotation or change the chart. Or it arrives as a second accent colour highlighting a different thing from the first. That's the focal-point failure, and it comes up later.

A chart that fails the test isn't always wrong. Sometimes it's a table drawn as a chart because the author felt a chart was expected. Sahyadri Diagnostics, another synthetic firm, publishes a monthly ops report with revenue, volume, turnaround-time breaches, cost per test and repeat share for eight centres. That's 40 numbers, and a reader looks things up in it. A table is the right form, sorted by whatever the reader looks up most. A chart of it would carry no message at all, because the author hasn't decided what the message is.

Writing the message before the chart exists makes the test cheap, and that's section 4's subject. Running the test on other people's charts makes it stick. There the second message is always obvious, and you can see how invisible it was to the person who drew it.

2. The action title: a sentence with a verb and a number

An action title is the message written above the exhibit as a sentence. Three properties make one work, and someone who has never seen your data can check all three.

It's a sentence, so it has a subject and a finite verb. "Contribution per order turned positive in Q7" has both. "Contribution margin by quarter" has neither. It can't be true or false, so nobody can argue with it, and nobody can argue from it either.

You can check it against the body directly below it. A reader who doubts the title should be able to look down, find the evidence, and either accept the claim or catch you out. That's a demanding standard. It rules out titles that are true but supported somewhere else, which feel harmless and are the commonest reason a deck's slides don't prove their own headings.

It carries the number where the data supports one. "Contribution per order turned positive in Q7 and reached ₹12.37 in Q8" is stronger than the same sentence without the figure. The figure invites a check, and a title that invites a check is one the author has already checked.

The verb comes from the message type, and picking it is most of the work. A comparison across categories takes a verb of ranking or size: leads, carries, accounts for. A change over ordered time takes a verb of movement: rose, fell, turned, stalled, doubled. A relationship takes a verb of association, and takes it carefully, because "drives" is a causal claim and "tracks" is not. A part-to-whole takes a verb of composition: makes up, has shifted to. A deviation takes a verb of distance: missed, ran ahead of, sits below.

How precise to make the number is a decision you take on purpose. Turant's Q8 figure is ₹12.3680 per order. Written as ₹12.37, it's accurate to the paisa. Written as ₹12, it carries a 2.98% error in a figure the reader may divide into something else, so the rounding isn't free. Written as ₹12.4, it's fine for a slide and awkward for a currency. The rule that survives contact with practice is to round to the precision at which the claim still holds, and no further. If you can't say what precision that is, you haven't worked out what the claim depends on.

Length has a physical ceiling. On a sixteen-by-nine slide, at the type size a title needs to be legible from the back of a room, roughly fourteen words fit on two lines. Compare four candidates for the Turant chart:

The fourth is the interesting failure. Everything in it is true, and it's the sentence the reader most needs, so you'll be tempted to shrink the type until it fits. What it really shows is an exhibit carrying two messages. The repair is two exhibits with one title each, in order, so the second follows the first.

One more property, and it's what separates a title that works in a deck from one that works only in the author's head. The title must survive being read alone, without the chart. Someone will forward your deck. Someone will read only the headings at speed. A title that means something only when you can see the picture below it has failed a test it's certain to be given.

3. Label titles diagnosed

"Revenue by region" is a filing label. It tells the reader where they are, the way a folder tab does. Folder tabs are useful when you're looking for something and useless when someone's telling you something. Four kinds of label title turn up, each fails for its own reason, and each has its own repair.

The filing label names the contents: "Q2 results", "Marketing update", "Financials". It comes from an author organising their material when they should be organising their argument, and the repair is to ask what the reader should conclude from these contents. The dimension label names the cut: "Revenue by region", "Margin by quarter", "Tests by centre". It's the most common of the four and the most seductive, because it describes the chart accurately and so feels complete. The topic label names a subject with no cut at all: "Unit economics", "Customer retention". You'll usually find it above a chart that has three messages and couldn't choose. The question label poses a question the slide then answers: "What is driving margin?" It's closer to useful than the others, since a question at least implies an answer exists, but it still makes the reader do the work of finding it.

You can repair a dimension label mechanically, and it's worth practising until it's fast. Take "Contribution margin by city". Ask what the cut shows. Sort the cities by the latest value, and the answer is that seven of the eight rows are now positive. One, Chennai, is at −₹9.75 per order against a network figure of ₹12.37, a gap of ₹22.12. Now ask whether that's the message worth carrying. If it is, the title writes itself: "Chennai is the only city still losing money on every order, at −₹9.75 against a network ₹12.37." The chart may need to change to support it, which is why you do the title first.

A label title is correct in two places, and only two. The first is a reference exhibit in an appendix whose job is lookup. There "Contribution margin by city, Q3 to Q8" is the right title, because the reader arrived looking for a city. The second is a standing table in an operating pack that the same people read every month. The label works as a stable address there, and the message goes in the annotation. Anywhere else, a label title on a page someone's making a decision from is a slide that hasn't finished being written.

There's one more thing a label title hides, and it's why experienced reviewers go after them first. A label title fits any data. Change every number on the chart and "Revenue by region" is still accurate. So the title tells you nothing about what the author found. In a deck whose titles are all labels, no reader can tell from the headings alone whether the analysis found anything at all.

4. Writing the title first, and letting it choose the chart

The ordinary sequence is data, then chart, then title. The chart gets built because the data's available, and the title gets written last, when the author is tired and the chart's already made. It's the moment a dimension label is easiest to type.

Turn it round. Write the claim first, as a sentence, from what you already believe the analysis shows. Then ask a single question: what would a reader need to see to believe this? The answer is the exhibit's specification, and the message type picks the form for you, so habit doesn't get a vote.

Watch it work on Turant. The claim is "Contribution per order turned positive in Q7 and reached ₹12.37 in Q8." The message type is a change over ordered time, so the form is a line. The axis carries rupees per order, and the horizontal runs Q1 to Q8 in order. Zero is on the axis and marked, because the crossing is the whole claim, and the Q8 point is labelled directly with its value. That specification took one sentence to produce, and it rules out several things the author might otherwise have added: a second series, a stacked breakdown of the components, and a projection.

Now change the claim and watch the exhibit change with it. Say the claim is "Three fifths of the ₹50.58 swing came from gross margin and two fifths from lower variable cost". Now the message type is composition of a change, the form is a two-bar decomposition or a small waterfall, and the eight quarters vanish because the claim isn't about them. The numbers are ₹31.28 from gross margin, which is 61.84% of the swing, and ₹19.30 from variable cost, which is 38.16%. They sum to ₹50.58 exactly, and that's your check that the decomposition is complete.

If the claim is "Breakeven needs 2.23 times the current order volume", the message type is a comparison of two quantities and the form is two bars, or a single bar with a target line. The units change too, from rupees per order to crore orders. Same company, same quarter, three claims, three different exhibits, and none of them is a redraw of another.

This prevents a familiar failure. An author builds an exhibit, can't say what it shows, and reaches for a dimension label to get past the problem. If you can't write the title before you build, you don't yet know what you're building, and fifteen minutes of not knowing turns into an hour of drawing.

One caution about writing the title first, because the habit has a real way of going wrong. Decide the claim before you've run the numbers and the deck ends up asserting something the data doesn't support. The order that works is: know your findings, write the claim from them, build the exhibit that tests the claim, and then be ready to change the claim when the exhibit contradicts it. That last step is the one people skip. An exhibit that disagrees with its title is telling you something, and a softer title is the wrong response.

5. The title that overstates, and the three tests before a claim ships

A claim in a title is a claim you'll be asked to defend, usually by the one person in the room who knows the number. Three tests, run in order, catch almost everything.

The first test is recomputation, and it's stronger than checking. Recompute from the source, in a way that doesn't reuse the intermediate you already have. Turant's breakeven multiple is fixed costs over contribution per order, divided by current volume. Dividing 44.60 by 12.3680 gives 3.6061 crore orders, and 3.6061 over 1.62 is 2.2260. If the title says 2.2 times, it's right. If it says "nearly three times", it's wrong by a margin anyone with a calculator will find in eleven seconds.

The second test is scope. A title makes a claim about a period, an entity and a unit, and all three must fall inside what the chart contains. "Turant's unit economics turned positive" is a claim about the company. The chart shows contribution margin per order, which leaves out fixed costs entirely, so the claim reaches wider than the evidence. The narrower true version is "contribution per order turned positive". The difference between those two sentences is ₹24.56 crore of quarterly EBITDA loss.

The third test is strength. Three strength errors turn up often enough to be worth naming.

The first is the causal upgrade. A scatter of fourteen US industrials shows return on invested capital and enterprise value over invested capital correlating at 0.83, with an R-squared of 0.69. "Higher returns drive higher multiples" is a causal claim from cross-sectional data, and the chart doesn't support it. What the picture shows is "companies earning higher returns on capital trade at higher multiples of it".

The second is the level-rate confusion. Turant's EBITDA loss narrowed from ₹71.06 crore to ₹24.56 crore across eight quarters, and every one of those quarters was a loss. "The quarterly loss has narrowed by ₹46.50 crore" is true, but it's a level. As a rate, the narrowing across seven steps runs ₹6.64 crore a quarter. "The company is close to profitability" is a different claim. It needs the volume arithmetic, and the volume arithmetic says 2.23 times current orders.

The third is the one that catches good analysts: the sign-change trap. Contribution per order went from −₹2.14 in Q6 to ₹12.37 in Q8. The ratio of those two numbers is −5.79. The percentage change computed against the absolute value of the base is 679.30%. A spreadsheet will happily produce both, and neither means anything. A percentage change measures movement relative to a base, and this base is on the other side of zero. So state it in absolute terms: contribution per order improved by ₹14.50 across those two quarters, having crossed zero in between. The same trap fires on any series that crosses zero, which includes free cash flow, operating profit at a loss-making business, net migration, and the change in an inventory position.

A fourth strength error deserves a mention, because it's a rounding decision dressed up as a claim. Sahyadri's network turnaround-time breach rate rose from 4.12% to 5.71%. That's 1.59 percentage points, or a ratio of 1.39 times. "Breaches doubled" is false. "Breaches rose by more than a third" is true and duller, and duller-and-true is the only version that survives the meeting.

All three tests rest on the recomputation habit from the course's quantitative work, applied here to prose instead of to a model. Nothing goes into a title that you haven't computed yourself, in a form you could show. It costs a few minutes per exhibit, and it's what separates a deck that gets believed from one that gets audited.

11. Sequencing exhibits so each one advances the case

A single good exhibit is a component. The sequence is the argument, and its test is stricter than the test for its parts. Each exhibit must add something the previous one didn't, and the conclusion must need that addition.

Read the titles in order, ignoring the exhibits, and check three things at each step. The title has to say something new. It has to follow from what came before, so that a reader who accepted the previous claim is ready for this one. And it has to be needed: take it out and the conclusion gets weaker.

Turant's board sequence runs five exhibits, and each answers the question the one before it provokes.

The sequence has a shape, and it's the shape most analytical arguments take. Exhibit 1 establishes the fact. Exhibit 2 tests whether the fact is real. Exhibit 3 complicates it, and that's the move that makes the reader trust the deck, because a sequence with no complication reads as advocacy. Exhibit 4 sizes the remaining problem. Exhibit 5 settles what to do about the size.

The commonest sequencing fault is an exhibit whose title is implied by the one above it. If exhibit 3 said "EBITDA is still negative" and exhibit 4 said "The company is not yet profitable", the second adds nothing. Delete it, because rewording won't help. The second commonest is the exhibit that arrives out of order, leaving the reader to hold a claim in suspense until two slides later. You fix that one by moving it, and nothing needs cutting.

There's an arithmetic reason to care about a weak exhibit, beyond taste. Take a five-exhibit set where four score 0.90 on the seven-rule check and one scores 0.72. It averages 0.864 on the machine. A machine mean of 0.864 needs a rubric self-score of 0.954 to reach a 0.90 composite, where the four strong exhibits on their own would have needed 0.90. One weak exhibit costs the whole set. The cheapest repair is usually deletion, because an exhibit that's hard to get above 0.80 is often an exhibit with no message.

12. The appendix: what earns a page and what earns a backup

Every analysis produces more exhibits than the argument needs. Deciding what to do with each one takes a two-question test with three possible outcomes, and taste doesn't come into it.

Does this exhibit move the decision the deck exists to serve? If yes, it earns a page. If no, ask the second question: would a reader in the room plausibly ask for it? If yes, it earns an appendix slot. If no, it's cut, and the cut counts as a saving.

Run it on Turant's eleven candidates. Five move the decision and become the sequence above. Four are questions you can see coming, and they go to the appendix: contribution margin by city, rider utilisation by city, the category mix of the basket, and a competitor pricing index. Two answer nothing anyone will ask, orders by hour of day and app downloads by channel. Cutting them saves 50 minutes of the 345 the full set would have taken, or 14.49% of the build.

Two properties keep an appendix working, so it doesn't turn into a graveyard. It's indexed, so a page at the front lists what's back there and the presenter can reach any exhibit in one move. And every appendix exhibit still has an action title. The moment it's shown, it's doing the same job a page exhibit does, and a label title at that moment costs the presenter the twenty seconds of explanation the title should have carried.

The appendix also carries a promise. If the front of the deck asserts something and a reader asks how you got there, the derivation should be at the back. A summary that promises support the appendix doesn't contain is the fastest way to lose a room. The check is mechanical. For each claim in your five titles, name the appendix exhibit that supports it if challenged, and build the one you can't name.

13. The crown standard: what a chart satisfies before it ships

The seven machine-checked rules apply to every exhibit in this branch, and here all seven have content specific to a chart.

The composite is 0.6 times the machine score plus 0.4 times your own rubric self-score, which you commit before the machine's is revealed. The crown threshold is 0.90, against the branch's usual 0.80. Do that arithmetic once, because it changes what you work on. A machine score of 0.95 with a rubric of 0.60 blends to 0.810, which fails. A machine score of 0.86 with a rubric of 0.95 blends to 0.896, which also fails. To reach 0.90 with a rubric self-score of 0.85 you need a machine score of 0.9333, and with a rubric of 0.80 you need 0.9667.

There's a harder implication. If each of the seven rules independently had a 0.9 chance of passing, the chance of all seven passing on one exhibit would be 0.478, roughly a coin flip. In practice, exhibits don't fail independently. The rules are correlated because they come from one decision: knowing what the exhibit is for. An exhibit built from a written message before any drawing tends to pass all seven. One built from whatever data was available tends to fail three or four together.

The self-score is where you record the judgement the machine can't make, so commit it truthfully and specifically. Three questions carry most of the weight. The first asks whether the claim in your title is true and whether you recomputed the number in it. The second asks whether this is the message that matters to the decision or the one your data happened to make easy. The third asks whether a reader who disagrees with you could find, on this exhibit, the evidence that would change your mind.

{
  "skill": "Stress-testing an exhibit's action title against the exhibit's own data before it ships",
  "use": "After you have written the title yourself and recomputed its number, paste the chart's underlying data table and your candidate title into an assistant and ask it to attack the claim on scope, strength and arithmetic. Use it to find the overstatement you cannot see in your own sentence; never let it write the title, because the title is the claim you will defend in the room.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "Python or a spreadsheet for the independent recomputation", "The Slide Studio's linter for the seven-rule machine score"],
  "prompt": "Here is the data table behind one chart: <PASTE TABLE>. Here is my candidate title: <TITLE>. Do five things and nothing else. 1) Recompute every number in my title from the table and show the arithmetic. 2) State whether the claim's scope (period, entity, unit) is fully contained in this table, and name anything it asserts that the table cannot support. 3) Check the claim's strength: flag any causal verb used on cross-sectional or correlational data, any level-versus-rate confusion, and any percentage change computed across a sign change. 4) List every other claim this same chart could carry, so I can see whether I have smuggled in a second message. 5) Say which single element of the chart the title asks the reader to look at. Do not propose a replacement title.",
  "verify": "Recompute the title's number yourself from the source, by a route that does not reuse the intermediate the assistant used, and check it agrees to the precision the title states. Where the assistant flags a scope problem, resolve it against the chart's own axes and footnotes rather than against its explanation. Where it lists additional claims the chart could carry, check each against the exhibit by eye: any claim you can see in the picture but did not intend is a second message, and the fix is to split the exhibit or remove the element that carries it. An assistant that reports the claim is fine, with no arithmetic shown, has not run the check and its answer is not admissible.",
  "diy": "You must be able to run the three overstatement tests cold, on paper, in about two minutes an exhibit: recompute, check scope, check strength. That is what the crown gate scores and it is what a room tests when someone challenges a number. The sign-change trap in particular has to be automatic, because a spreadsheet will compute a percentage change across zero without complaint and the resulting figure looks entirely plausible on a slide.",
  "market": "IN"
}

Common mistakes & how experts think differently

1. Writing the title after the chart. The chart gets built because the data was available, and by the time a title is needed the author is describing what they drew. That produces a dimension label every time, because the plain description of a chart of margin by quarter is "margin by quarter". An expert writes the sentence first and makes the exhibit earn it. That also means they build fewer exhibits, because a claim they can't write is a chart they don't need.

2. Treating a percentage change across zero as a growth rate. A spreadsheet will divide −2.14 into 12.37 and return −5.79, or compute 679.30% against the absolute base, and neither number means anything. It's the most common numerical overstatement in decks about businesses that have just turned profitable. It's dangerous because the arithmetic is real and the output is plausible, and the error only shows up when someone asks what the base was. Experts state the swing in absolute terms and say that it crossed zero.

3. Confusing a unit-economics claim with a profitability claim. Contribution per order turning positive is a real event, and it isn't profitability. At Turant the distance between them is ₹24.56 crore of quarterly EBITDA loss and a breakeven volume 2.23 times current orders. This passes review so often because both claims are optimistic and the weaker one is easier to defend, so nobody in the room has an incentive to separate them. An expert states the narrower claim and puts the wider one in the next exhibit, where it can be sized.

4. Adding a second series because it is related. Revenue and retention are both about the same customers, so they seem to belong together. The moment they go on one chart with two axes, the author has picked a scaling that invents a story. Move a right-hand axis from 80 to 95 and a retention line crosses ARR between the fifth and sixth quarter. Set it to 86 to 94 and the same two series cross between the first and second. Set it to 0 to 100 and they cross between the seventh and eighth. The underlying series never moved. An expert either indexes both to a common base or, more often, builds two exhibits, because these were two separate messages all along.

5. Highlighting three things. Three accents is a common draft state, and it comes from an author with three findings and one slide. The reader has three pairwise comparisons to resolve before reading anything, and where the eye goes first becomes arbitrary. Experts pick the one element that carries the title's claim and grey everything else, and that usually shows the other two findings wanted exhibits of their own.

6. Giving each small-multiple panel its own axis. Most charting defaults do this, and the result is a grid where the flattest series looks the most dramatic. The nine-sector grid makes it concrete. Health care's 3.57-point range would fill a panel that a 29.73-point range fills elsewhere, an 8.32-fold exaggeration with nothing on the page to warn the reader. An expert fixes the axis across all panels before looking at the result. A panel that then looks empty is information, and they treat it that way instead of as a formatting problem.

7. Building a dashboard by accumulation. Every stakeholder asks for one more tile and nobody ever asks for one fewer. So a page that started as six tiles becomes fourteen, and the pairwise comparisons go from 15 to 91. You don't see the failure happening, because each individual request was reasonable. Experts run a fixed tile budget with an explicit swap: a new tile displaces a named old one, and if nobody will name the one to drop, the new tile wasn't important.

8. Keeping an exhibit because it took a long time to build. The forty minutes already spent are gone whether the exhibit ships or not. The only question is whether it moves the decision or answers a question the room will ask. Turant's eleven candidates included two that did neither. Cutting them saved 50 of 345 build minutes, which is 14.49% of the work, on exhibits that would have cost the reader attention as well. Experts make the cut list before they build anything, and that's what turns the saving from 14% into most of it.

9. Letting the title claim what the body cannot prove. A title saying the company is close to profitability, above a chart of contribution margin per order, is a title the body contradicts. The machine check catches the form of this failure and misses the substance. An expert reads the title and then asks, with a finger on the exhibit, where the proof is. If the finger has nowhere to go, either the chart changes or the title does.

10. Using a label title in a deck because a template used one. House templates often ship with placeholder titles of the dimension-label kind, and the placeholder becomes the convention. The cost is that a reader skimming headings learns nothing about what you found. The repair is one pass at the end: read every title alone, out of order, and rewrite any that could sit above a different dataset unchanged.

Worked examples

Every figure below comes from one computation, and the Verify block at the end of this section restates each one at full precision. All the companies are synthetic composites and labelled as such. The sector series are synthetic too, generated from a stated rule.

Worked example 1: Three candidate titles for one chart (India, ₹)

The setting. Turant (synthetic) runs a quick-commerce network in seven Indian cities plus nine smaller ones. Across eight quarters its average order value ran ₹412, ₹421, ₹433, ₹440, ₹448, ₹455, ₹461, ₹468, and its gross margin on that order value ran 12.4%, 13.1%, 14.0%, 14.8%, 15.6%, 16.3%, 17.0%, 17.6%. Variable cost per order, which is delivery, packaging and payment charges together, ran ₹89.30, ₹84.70, ₹80.10, ₹76.10, ₹78.30, ₹76.30, ₹73.80, ₹70.00. Quarterly order volume ran 0.86, 1.02, 1.15, 1.28, 1.39, 1.48, 1.55 and 1.62 crore orders, and quarterly fixed costs ran ₹38.20, ₹39.60, ₹41.10, ₹42.40, ₹43.10, ₹43.80, ₹44.20 and ₹44.60 crore.

Contribution per order is gross margin per order minus variable cost per order, so the eight quarters give:

The chart is a single line of contribution per order across the eight quarters, with zero marked. Three people propose three titles, so take them one at a time.

The wrong version, first candidate. "Contribution margin by quarter." It's a dimension label. It describes the cut and makes no claim. It would be just as accurate if every number on the chart were reversed in sign, and a reader skimming the deck's headings learns from it that Turant has quarters and a contribution margin. The four seconds they spent on it changed nothing about what they think.

The wrong version, second candidate. "Turant is now profitable and margins have tripled." Both halves fail, and both fail on the numbers. Taste doesn't come into it.

The first half fails on scope. The chart carries contribution per order, which is gross margin less variable cost, and contains no fixed costs at all. Turant's fixed costs in the eighth quarter were ₹44.60 crore against contribution of ₹20.04 crore, so quarterly EBITDA was −₹24.56 crore. All eight quarters were EBITDA-negative. The volume needed to cover fixed costs at the current contribution rate is 44.60 / 12.3680 = 3.6061 crore orders a quarter, which is 3.6061 / 1.62 = 2.2260 times the current 1.62 crore. The claim's scope is the company. The evidence's scope is one line of the unit economics. The gap between them is a factor of 2.23 in volume.

The second half fails on arithmetic. "Tripled" implies a ratio, and the two quarters someone would be comparing are Q6 at −₹2.14 and Q8 at ₹12.37. The ratio is 12.3680 / (−2.1350) = −5.7930, which is negative and meaningless. A percentage change against the absolute value of the base gives 679.30%. A spreadsheet will produce it without complaint, and it describes nothing, because a percentage change measures movement relative to a base and this base is on the other side of zero. The nearest defensible statement is absolute: contribution per order improved by ₹14.50 across those two quarters, crossing zero in between. If the intent was the last single quarter, 12.3680 / 4.5700 = 2.7063, so Q8 is 2.71 times Q7. That's true, and it's a one-quarter observation off a base of ₹4.57.

The corrected version. "Contribution per order turned positive in Q7 and reached ₹12.37 in Q8." It's twelve words, it fits two lines at slide type size, and you can check every element of it against the picture. The line crosses zero between the sixth and seventh points, and the eighth point carries a direct label reading ₹12.37.

Here's the exhibit that supports it, described as a layout. A single line, mid-grey from Q1 to Q6 and in the accent colour from Q6 to Q8, on a vertical axis in rupees per order running from −₹40 to ₹15 with a thin zero rule. The horizontal axis is labelled Q1 to Q8, the Q8 point carries a direct label, and one short annotation at the crossing reads "turned positive". No gridlines, no legend, no callout box, and no labels on the other six points. Elements competing for first attention: one.

Read it. The corrected title claims less than the second candidate, and it's the only one of the three a chief financial officer can't puncture. What it leaves out matters as much as what it says. It doesn't say the business is profitable, because that needs the volume exhibit. It doesn't say the improvement will continue, because eight points of history aren't a forecast. And it doesn't say why the improvement happened. That's a second message with its own exhibit: ₹31.28 of the ₹50.58 swing came from gross margin and ₹19.30 from lower variable cost, and those two sum to the swing exactly.

The checklist to apply. Before the title ships:

1. The title is a sentence with a finite verb.
2. Every number in it was recomputed from the source.
3. The claim's period, entity and unit all sit inside what the exhibit contains.
4. No ratio or percentage change is computed across a sign change.
5. A reader can point at the element of the chart that proves it.

Worked example 2: Nine crowded lines, rebuilt as small multiples (US, $)

The setting. Nine US sector indices (synthetic) are tracked monthly for twelve months from a common base of 100. The series are generated from a common market path of +1.8%, −0.9%, +2.4%, +1.1%, −1.6%, +2.0%, +0.7%, +1.3%, +2.9%, −3.4%, +4.1%, +2.6%, with a sector beta and a sector drift applied and a small idiosyncratic term whose spread widens after the eighth month. Full-year returns come out as:

The wrong version. All nine on one line chart with a nine-entry legend down the right side. The chart is fairly scaled, correctly formed for a change-over-time message, and unreadable. Nine series make 36 pairs, and across the twelve months those pairs cross 31 times, an average of 0.86 crossings per pair. Consumer discretionary and materials cross each other six times on their own. To follow one sector, a reader first has to find it in the legend, match a colour, and then track it through the crossings. Nine colours also break the accent rule from the declutter work. Every series is emphasised, so none is.

The title on the crowded version reads "Sector performance, monthly index". That's a dimension label, and it's a fair description of what the author did: they plotted what they had.

The corrected version. A three-by-three grid of small panels, one sector per panel, ordered by full-year return, so the grid runs from communication services in the top-left to health care in the bottom-right. Every panel shares one vertical axis covering the whole dataset, 96.57 to 126.30, a span of 29.73 index points. Each panel shows all nine series in a very light grey and its own sector in the accent colour, so every small chart is one sector against the bundle. Each panel is labelled with the sector name and its full-year return, and the axis labels appear once, on the left column and bottom row.

The title becomes a claim: "Eight of the nine sectors moved together all year; health care moved against them and was the only one to end lower."

The claim survives all three overstatement tests. The number recomputes: health care ends at 98.15, a return of −1.85%, and it's the only negative in the set. The scope matches, because the claim is about these nine series over these twelve months and nothing wider. The strength is right, because "moved together" is an association claim and the association has been measured. Correlate each sector's monthly returns against the equal-weighted average of the other eight and you get +0.98 for consumer discretionary, +0.96 for communication services, +0.92 for industrials, +0.89 for technology, +0.88 for both financials and materials, +0.77 for consumer staples, +0.68 for utilities, and −0.84 for health care.

The shared-scale argument, computed. Software defaults will give each panel its own axis, and here's what that costs. Health care's own range across the year is 3.57 index points against the shared span of 29.73. Under free scaling its 3.57-point wiggle would fill the same panel height that a 29.73-point move fills elsewhere, an exaggeration of 29.7320 / 3.5742 = 8.32 times. Utilities, at a 4.99-point range, would be exaggerated by 29.7320 / 4.9884 = 5.96 times. Read a free-scaled grid and you'd conclude that health care and utilities were the year's most eventful sectors, which is the exact opposite of the truth. Nothing on the page would warn you.

Read it. The rebuild doesn't add information. It removes a task. The crowded chart contains the health care fact, and a determined reader with two minutes could dig it out. The grid delivers it in about two seconds, because the comparison people make fastest is shape against shape. Following lines is slower. What the rebuild can't do is tell you why health care ran the other way. The fair treatment of that on the slide is a footnote saying the data doesn't answer it.

The checklist to apply.

1. Count the series. Above five on one chart, price the crossings before deciding.
2. If panels, one shared vertical scale unless the units really differ.
3. Order the panels by the value the message is about.
4. Ghost the full set behind each panel so the comparison survives the split.
5. Write the title as the claim the grid makes visible, not as a description of the grid.

Worked example 3: One chart carrying two messages (US, $)

The setting. Larkspur Systems (synthetic) is a US software company. Ending annual recurring revenue across eight quarters is $41.2m, $46.8m, $52.9m, $58.4m, $63.1m, $67.2m, $70.6m and $73.9m. Annualised gross retention across the same quarters is 86.4%, 87.1%, 88.0%, 89.2%, 90.1%, 91.4%, 92.3% and 93.1%. Net additions are $5.6m, $6.1m, $5.5m, $4.7m, $4.1m, $3.4m and $3.3m from the second quarter onward.

Two facts are true and both matter. Year-on-year ARR growth decelerated from 53.16% in the fifth quarter to 26.54% in the eighth, a fall of 26.61 points. Gross retention improved by 6.7 points over the same eight quarters, from 86.4% to 93.1%.

The wrong version. One chart, ARR as columns on a left axis in dollars running $40m to $75m, retention as a line on a right axis in percent, titled "ARR and retention". Two units, two axes, one picture, and a title that is a topic label because the author could not choose between the two messages.

Here's the specific damage a dual axis does: the author picks the right-hand scale, and the scale picks the story. Hold the data and the left axis fixed and vary only the right axis:

Three different stories, one dataset, and nothing on the chart tells a reader that a choice was made. The crossing point is an artefact of the author's axis, and the reader has no way to see it.

The corrected version. Two exhibits, sequenced, each with one message and one focal point.

The first is a column chart of net ARR additions by quarter, $5.6m, $6.1m, $5.5m, $4.7m, $4.1m, $3.4m, $3.3m, with the Q3 peak and the Q8 value labelled and the declining run in the accent colour. Its title: "Net ARR additions peaked at $6.1m in Q3 and have fallen every quarter since, to $3.3m." The reader sees the deceleration directly in the heights and doesn't have to infer it from a rising cumulative line. That's the second reason to prefer this exhibit over an ARR level chart: a level chart of ARR rises in every quarter and looks like unbroken success.

The second is a line chart of gross retention, 86.4% to 93.1%, on an axis running from 85% to 95%, with the endpoints labelled. Its title: "Gross retention improved 6.7 points to 93.1%, so the deceleration is coming from new business rather than churn."

That second title does the sequencing work described earlier. It states its own fact and connects it to the previous exhibit's fact, and that connection is what turns two charts into an argument.

The check that the split is honest. If retention improved and growth still decelerated, the arithmetic has to land somewhere, and it lands in gross new business. ARR grew 79.37% across the eight quarters while net additions fell 45.90% from their peak, and both statements come from the same table. A reader who wants to challenge the second title can ask for the gross-new-versus-churn split, and that's the exhibit that belongs in the appendix.

Read it. A better colour palette wouldn't fix the dual-axis version, because the problem isn't formatting. It's a symptom of an author holding two messages and refusing to spend a second slide. That refusal has a cost. The one slide makes neither point cleanly, and it invites a question about the axis that the author can't answer well.

The checklist to apply.

1. Say the chart's message in one clause. If it needs "and", split it.
2. Two units on one plot is a split, an index, or a small multiple, never two axes.
3. Where a rate is the message, plot the rate, not the level it accumulates into.
4. Give the second exhibit a title that connects to the first.
5. Name the exhibit that would answer the obvious challenge, and put it in the appendix.

Worked example 4: A wall becomes a dashboard (India, ₹)

The setting. Sahyadri Diagnostics (synthetic) runs eight pathology centres in Maharashtra. In the reporting month the centres did 41,200, 38,600, 22,400, 19,800, 14,100, 11,900, 9,700 and 8,300 tests, for a network total of 166,000. Revenue by centre was ₹128.40, ₹121.90, ₹66.80, ₹58.10, ₹40.30, ₹33.20, ₹26.40 and ₹22.10 lakh, totalling ₹497.20 lakh, which is ₹299.52 a test across the network. Turnaround-time breaches, meaning tests delivered later than the promised time, ran 9.1%, 7.8%, 3.2%, 2.9%, 3.4%, 2.6%, 3.9% and 3.1% of each centre's volume. The prior month's network breach rate was 4.12%.

The volume-weighted network breach rate is Σ(volume × breach) / 166,000 = 5.71%, up 1.59 percentage points from 4.12%, a ratio of 1.39 times.

The wrong version. A monthly operations page with fourteen tiles: revenue, revenue growth, test volume, volume growth, revenue per test, cost per test, gross margin, turnaround-time breaches, repeat-customer share, home-collection share, average collection days, staff utilisation, equipment downtime, and net promoter score. Every tile is a number with a small arrow. The page is titled "Operations dashboard, September".

Two things are wrong, and neither is about any one tile. Fourteen tiles admit 91 pairwise comparisons, and a reader trying to work out what changed gets no guidance about which ones matter, so they make whichever catch their eye. At four seconds a tile, the page costs 56 seconds to scan before any thinking starts. And there's no claim anywhere on it, so ten readers form ten different views of the month.

The corrected version. Six tiles, a fixed reading order, and a dated subtitle that carries the month's message.

The standing title stays "Operations dashboard", because the same people read the page every month and a stable address is worth something. The subtitle changes each month and carries the claim: "Turnaround breaches rose to 5.71% from 4.12%; 48% of network volume sits in the two centres driving it."

The top row is the argument. The bottom row reassures the reader that nothing else moved. The reading order is designed on purpose: the reader enters top-left and gets the claim, moves right for the cause, moves right again for the exposure, and drops to a bottom row that says the rest of the business is where it was.

The arithmetic that justifies the cut. Six tiles admit 15 pairwise comparisons against 91, a reduction of 83.52%, and the scan cost falls from 56 seconds to 24. The eight tiles removed weren't wrong. They answered questions nobody was asking this month, and a swap rule governs their return: a new tile displaces a named old one.

The independent check. You can verify the breach claim from the other direction. Take the two Pune centres out, and the remaining six did 86,200 tests at a volume-weighted breach rate of 3.1501%, against 8.4712% for the 79,800 tests the two Pune centres did. Rebuild the network from those two groups and you get (79,800 × 8.4712 + 86,200 × 3.1501) / 166,000 = 5.7081%, which is the figure in the top-left tile, and 8.4712 / 3.1501 = 2.6892 is the ratio in the second.

Read it. The six-tile page is a different artefact from the fourteen-tile page, built from a decision about what the page is for, so calling it a summary misses the point. The eight removed tiles still exist in the monthly detail pack, where anyone who wants staff utilisation can find it. What the dashboard buys you is that every reader leaves with the same view of the month, and that's the only thing a standing page is actually for.

The checklist to apply.

1. Name the repeated question the page answers, as a single written sentence, before choosing any tiles.
2. Put the claim in the top-left position and date it.
3. Every tile carries a metric, a comparison and a status.
4. Fix the tile budget and enforce a swap rule.
5. State the refresh time on the page.

Worked example 5: The exhibit inside a document, and the caption that carries it (US, $)

The setting. An analyst's written note covers fourteen synthetic US industrial companies. For each, the note carries return on invested capital for the latest fiscal year and enterprise value as a multiple of invested capital, at a common valuation date. The cost of capital assumed throughout is 9.5%.

The wrong version. A scatter of the fourteen points with a fitted line, captioned "Exhibit 4. ROIC versus EV/IC". Two failures, and the second is the expensive one. The caption is a dimension label, so a reader who skims captions learns that the note plotted two variables against each other. And the exhibit is in the document because the analyst found the relationship interesting, when it should be there because a specific claim in the prose needs it. That shows up later, when nothing in the text refers to Exhibit 4 by number.

The corrected version. The same scatter, with three changes. Meridian Fluid is the one point in the accent colour, and it's labelled. The fitted line is a thin grey rule in place of a heavy trend line. And the caption does the work.

The caption: "Exhibit 4. Companies earning higher returns on capital trade at higher multiples of it; Meridian Fluid, at 22.0% return and 1.10x invested capital, trades 57.2% below the fitted relationship. Source: synthetic dataset, latest reported fiscal year, valuation date as stated in the note."

Here are the numbers behind the caption. Fitting a least-squares line through all fourteen points gives a slope of 0.117853 turns of invested capital per percentage point of return and an intercept of −0.023296, with an R-squared of 0.688714 and a correlation of 0.829888. Meridian's fitted value at 22.0% return is 2.5695x against an actual 1.10x, a residual of −1.4695 turns, which is 1.10 / 2.5695 − 1 = −57.19%. No other company is more than 0.44 turns from the line.

The independent check that makes the exhibit worth its page. Refit the line without Meridian. The slope becomes 0.144519, the intercept −0.313945, and the R-squared rises from 0.688714 to 0.993231, an increase of 0.304517. Thirteen of the fourteen companies sit almost exactly on a line and one doesn't. That's a fact about the shape of the data, and a correlation coefficient reported in a sentence would have hidden it completely. A correlation of 0.83 sounds like a moderately noisy relationship, when what you've got is a near-perfect one with a single large exception.

The fitted line also gives the note a reference point to use elsewhere. It crosses 1.0x invested capital at a return of 8.6828%, close to the 9.5% cost of capital assumed throughout, which is the relationship a reader trained in returns and value expects to see. Saying that in the caption would add a second message, so it goes in the prose that refers to the exhibit.

The verb, chosen carefully. The caption says companies earning higher returns "trade at" higher multiples. It doesn't say returns "drive" multiples. Fourteen companies at one date establish an association and nothing about direction or cause. And an R-squared of 0.69 across the full set isn't a strong enough relationship to carry a causal verb, even if the design allowed one.

Read it. This example is really about the caption-only test. A reader who reads only the note's captions, in order, should be able to rebuild the argument. "Exhibit 4. ROIC versus EV/IC" adds nothing to that, while the corrected caption adds a claim and a named exception. In a document the caption is the title, and it also has to survive being read without the paragraph around it.

The checklist to apply.

1. The exhibit exists because a specific sentence in the prose needs it, and that sentence refers to it by number.
2. The caption states the claim, names the exception, and carries the source and date.
3. The verb matches what the data design can support.
4. One accented point, chosen because it is the exception the claim names.
5. The exhibit appears after the first reference to it, never before.

Worked example 6: Five exhibits in sequence, and the appendix decision (India, ₹)

The setting. Turant's board meets to decide whether to approve the next funding round now or in two quarters. The analysis has produced eleven candidate exhibits. Cash at quarter end is ₹186 crore, capital spending in the quarter was ₹6.4 crore, and the quarter's EBITDA was −₹24.56 crore.

The wrong version. All eleven exhibits, in the order they were built, titled with dimension labels, running to eleven slides. The board spends its time on whichever chart is most colourful and reaches the funding question with nine minutes left. The decision gets deferred, because nobody has shown them the size of the remaining gap.

The corrected version. Two questions decide each exhibit. Does it move the decision? If not, would somebody in the room ask for it? That gives three groups: five exhibits earn a page, four earn an appendix slot, and two are cut.

The five that earn a page, in order, each answering the question the previous one provokes:

Exhibit 5's arithmetic: the quarter's cash burn is the EBITDA loss plus capital spending, 24.5638 + 6.4 = ₹30.9638 crore, which is ₹10.3213 crore a month, so 186.0 / 10.3213 = 18.02 months, or 6.01 quarters.

The four that earn an appendix slot are contribution margin by city, rider utilisation by city, the category mix of the basket, and a competitor pricing index. Each answers a question a director plausibly asks and none of them changes the funding decision.

The two that are cut are orders by hour of day and app downloads by channel. Neither moves the decision, and neither answers a question the room will ask. Cutting them saves 50 of the 345 minutes the full eleven would have taken, which is 14.49% of the build.

The sequence test, run on the titles alone. Read the five titles with the exhibits hidden. Each says something the one before it didn't: the turn, the reason for the turn, the fact that the turn hasn't reached the bottom line, the size of what remains, and the time available to close it. Nothing in title 4 is implied by title 3, and a reader who accepted 3 is ready for 4. Delete any one of the five and the argument has a hole. Without 2, the turn could be a one-quarter cost cut. Without 3, the deck reads as advocacy. Without 4, the gap is unsized. Without 5, the board has no basis for choosing between now and two quarters out.

What the sequence costs if one exhibit is weak. Say four of the five score 0.90 on the seven-rule check and the fifth scores 0.72. The set averages (0.90 × 4 + 0.72) / 5 = 0.864 on the machine. Reaching a 0.90 composite from there takes a rubric of (0.90 − 0.6 × 0.864) / 0.4 = 0.954, against the 0.90 the four strong exhibits would have needed on their own. The cheapest repair on a weak exhibit is usually deletion, and the reason is diagnostic. An exhibit that won't come above 0.80 is usually an exhibit whose message was never decided.

Read it. People get the appendix wrong in the other direction, by treating it as a place to put everything that didn't fit. It works when every appendix exhibit still carries an action title and the front of the deck has an index to them, so a director asking about cities gets an answer in one move, without a search. The promise runs the other way too. For each of the five page titles, name the appendix exhibit that would support it under challenge, and build the one you can't name.

The checklist to apply.

1. Two questions per candidate: does it move the decision, and would somebody ask for it.
2. The page set is ordered so each title answers the previous title's question.
3. No title is implied by the one above it.
4. Every appendix exhibit has an action title and an index entry.
5. Every page claim has a named appendix exhibit behind it.

Worked example 7: The sparkline column that changed the ranking (India, ₹)

The setting. Turant reports contribution margin per order by city for six quarters, Q3 to Q8. Each city's figure is the network figure plus that city's gap to the network. The gaps are re-centred each quarter so the volume-weighted gap is exactly zero, which makes the table reconcile to the company line by construction. Order shares by quarter for the seven named cities plus the nine smaller ones sum to 1.000 in every column.

The reconciliation holds in every quarter: the share-weighted city figures reproduce the network line to the sixth decimal place.

The wrong version. The table with one number per city, the latest quarter only, sorted descending, titled "Contribution margin by city". A reader gets the level ranking: Mumbai ₹17.25, Ahmedabad ₹16.05, Bengaluru ₹15.55, Hyderabad ₹13.95, Pune ₹12.75, Delhi NCR ₹12.15 and Chennai −₹9.75. From that page the obvious conclusions are that Mumbai is the best city and Delhi is nearly the worst.

The corrected version. The same table, with a sparkline column between the city name and the latest value. One shared scale runs across the whole column, from −₹29.47 to ₹17.25, every spark has a light zero rule, and the last value is labelled numerically beside each spark. The title becomes a claim: "Six of the seven named cities have crossed into positive contribution; Chennai has not, and Mumbai's lead is being competed away."

Two rows move a long way once you can see slope. Rank the seven named cities by their gain over the last two quarters and you get Ahmedabad ₹17.00, Delhi NCR ₹15.00, Hyderabad ₹14.20, Bengaluru ₹14.00, Mumbai ₹13.50, Pune ₹9.70, Chennai ₹7.50. Mumbai is first on level and fifth on slope. Delhi NCR is sixth on level and second on slope. Pune looks respectable at fifth on level, but it's sixth on slope. Its quarter-by-quarter gains of ₹7.46, ₹1.23, ₹4.53, ₹4.49 and ₹5.22 are the weakest of the mainland set after Chennai.

Why the shared scale matters here too. The seven cities' own ranges run from ₹14.92 for Chennai to ₹37.92 for Ahmedabad, against a shared column span of ₹46.72. Give each spark its own scale and Chennai, the one city that hasn't turned, would show the same visual amplitude as Ahmedabad, the fastest improver. That's an exaggeration of 46.72 / 14.92 = 3.13 times. A sparkline column has no axis labels at all, so free scaling there misleads the reader and they can't even detect it.

The number the table produces that no city row shows. Chennai carries 7.9% of orders at −₹9.75 against a network ₹12.37, so its weighted deviation from the network figure is 0.079 × (−9.75 − 12.37) = −₹1.75 per order. Take Chennai out and the remaining 92.1% of orders carry the average alone, so the network reads ₹14.27, an uplift of 1.7475 / 0.921 = ₹1.90. That second figure is the one a board will ask for. Put it in the sentence beneath the table, and don't add an eighth column for it.

Read it. The sparkline doesn't replace the number, and it never should. It adds the one dimension a level column can't carry, which is direction, and it does so in the width of a word. What it can't do is tell you the size of any point except the labelled one. A reader who needs the Q5 figure for Hyderabad still needs the table, so the sparkline column goes inside a table and doesn't replace it.

The checklist to apply.

1. One scale across the whole sparkline column, always.
2. Label one endpoint numerically so the shape has an anchor.
3. Mark the threshold that matters, here zero.
4. Sort by whichever of level or slope the title's claim is about, and say which.
5. Keep the numbers. A sparkline adds to a table and doesn't replace one.

Verify

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

Turant contribution per order: gross margin per order 412 × 0.124 = 51.088, 421 × 0.131 = 55.151, 433 × 0.140 = 60.620, 440 × 0.148 = 65.120, 448 × 0.156 = 69.888, 455 × 0.163 = 74.165, 461 × 0.170 = 78.370, 468 × 0.176 = 82.368. Contribution per order −38.212, −29.549, −19.480, −10.980, −8.412, −2.135, 4.570, 12.368. As a share of order value −9.2748%, −7.0188%, −4.4988%, −2.4955%, −1.8777%, −0.4692%, 0.9913%, 2.6427%. Contribution in ₹ crore −32.8623, −30.1400, −22.4020, −14.0544, −11.6927, −3.1598, 7.0835, 20.0362. EBITDA in ₹ crore −71.0623, −69.7400, −63.5020, −56.4544, −54.7927, −46.9598, −37.1165, −24.5638, all eight negative, narrowing by 46.4985 across the eight quarters, which is 46.4985 / 7 = 6.6426 a quarter. Order value times volume gives gross merchandise value 354.32, 429.42, 497.95, 563.20, 622.72, 673.40, 714.55, 758.16 ₹ crore, so Q8 EBITDA is −3.2399% of it.

Turant swing 12.368 − (−38.212) = 50.580, of which gross margin 82.368 − 51.088 = 31.280 (61.8426%) and variable cost 89.30 − 70.00 = 19.300 (38.1574%); the two sum to 50.580 exactly. Average improvement 50.580 / 7 = 7.225714 a quarter. Q7 to Q8 ratio 12.3680 / 4.5700 = 2.706346. Q6 to Q8 ratio 12.3680 / (−2.1350) = −5.792974; against the absolute base 679.2974%; absolute improvement 14.503. Breakeven 44.60 / 12.3680 = 3.606080 crore orders, 3.606080 / 1.62 = 2.225975 times. Rounding: 12.3680 to the rupee is 12, a 2.9754% error. Full data span 50.580, with zero sitting 75.5476% of the way up.

Nine sectors: full-year returns 26.2999, 23.1295, 11.4976, 10.9616, 10.7893, 9.7482, 9.4043, 4.9884, −1.8544 percent, spread 28.1542 points, index at month 12 126.2999, 123.1295, 111.4976, 110.9616, 110.7893, 109.7482, 109.4043, 104.9884, 98.1456. Crossings across 36 pairs 31, mean 0.8611, busiest pair 6. Dispersion by month 0, 3.33, 2.05, 5.86, 7.90, 5.48, 9.63, 11.24, 13.22, 19.99, 14.98, 23.37, 28.15, so 53.0277% of the year's dispersion arrives after month 8. Correlations with the average of the other eight 0.9761, 0.9555, 0.9152, 0.8863, 0.8816, 0.8812, 0.7664, 0.6817, −0.8372. Shared panel scale 96.5678 to 126.2999, span 29.7320; own ranges 26.2999, 23.1295, 11.4976, 11.0921, 10.9616, 10.1888, 9.5365, 4.9884, 3.5742; free-scale exaggeration 29.7320 / 3.5742 = 8.3186 for health care and 29.7320 / 4.9884 = 5.9603 for utilities.

Larkspur: ARR 41.2, 46.8, 52.9, 58.4, 63.1, 67.2, 70.6, 73.9; year-on-year growth from the fifth quarter 53.1553%, 43.5897%, 33.4594%, 26.5411%, a deceleration of 26.6142 points; net additions 5.6, 6.1, 5.5, 4.7, 4.1, 3.4, 3.3, a fall of 45.9016% from the peak; ARR growth over eight quarters 79.3689%; gross retention 86.4 to 93.1, +6.7 points. Dual-axis crossings, with the left axis held at 40 to 75: right axis 80 to 95 crosses between Q5 and Q6; 86 to 94 between Q1 and Q2; 0 to 100 between Q7 and Q8.

Sahyadri: volumes 41,200, 38,600, 22,400, 19,800, 14,100, 11,900, 9,700, 8,300, total 166,000; revenue 128.40, 121.90, 66.80, 58.10, 40.30, 33.20, 26.40, 22.10 ₹ lakh, total 497.20, ₹299.5181 a test. Volume-weighted breach 5.7081% against 4.12% prior, +1.5881 points, ratio 1.3855. Two Pune centres 79,800 tests at 8.4712%; other six 86,200 at 3.1501%; ratio 2.6892; reconstituted (79,800 × 8.4712 + 86,200 × 3.1501) / 166,000 = 5.7081%. Volume share 48.0723%, revenue share 50.3419%. Tiles: 14 × 13 / 2 = 91 comparisons, 6 × 5 / 2 = 15, a reduction of 83.5165%; at 4 seconds a tile, 56 seconds against 24.

Fourteen industrials: slope 0.117853, intercept −0.023296, R-squared 0.688714, correlation 0.829888. Meridian fitted 2.5695, actual 1.10, residual −1.4695, discount −57.1895%; next largest residual +0.4341 (Norwood). Refit excluding Meridian: slope 0.144519, intercept −0.313945, R-squared 0.993231, up 0.304517. The fit crosses 1.0x at a return of 8.6828%, and at a 9.5% cost of capital it sits at 1.0963x.

Turant board pack: five pages, four appendix, two cut; build minutes 125, 170 and 50 against 345 for all eleven, so the cut saves 14.4928%. Quarterly cash burn 24.5638 + 6.4 = 30.9638 ₹ crore, monthly 10.3213, runway 186.0 / 10.3213 = 18.0210 months and 6.0070 quarters. A five-exhibit set with four at 0.90 and one at 0.72 averages 0.864.

Turant cities, Q8: 17.2473, 16.0473, 15.5473, 13.9473, 12.7473, 12.1473, −9.7527 for the seven named and 12.3473 for the aggregate row; share-weighted across all eight rows the six quarters reproduce −19.480000, −10.980000, −8.412000, −2.135000, 4.570000, 12.368000. Last-two-quarter gains 17.00, 15.00, 14.20, 14.00, 13.50, 9.70, 7.50. Pune's quarterly gains 7.4584, 1.2324, 4.5250, 4.4871, 5.2169. Own ranges 29.32, 31.02, 32.22, 30.72, 22.92, 14.92, 37.92, shared span 46.72, free-scale exaggeration for Chennai 46.72 / 14.92 = 3.1314. Chennai drag 0.079 × (−9.7527 − 12.3680) = −1.7475; network excluding Chennai 14.2654, an uplift of 1.8974.

Crown arithmetic: 0.6(0.95) + 0.4(0.60) = 0.8100; 0.6(0.86) + 0.4(0.95) = 0.8960; 0.6(0.90) + 0.4(0.90) = 0.9000; (0.90 − 0.4 × 0.85)/0.6 = 0.9333; (0.90 − 0.4 × 0.80)/0.6 = 0.9667; (0.90 − 0.6 × 0.95)/0.4 = 0.8250; (0.90 − 0.6 × 0.864)/0.4 = 0.9540; (0.90 − 0.6 × 0.90)/0.4 = 0.9000; 0.9^7 = 0.478297. Focal points: 3 accents give 3 pairwise comparisons, 5 give 10, 6 give 15.

Practice set

Work each one before you read the solution. A calculator is assumed. Numeric answers within 2% of the stated figure count as correct unless a problem says otherwise. Where a problem asks for a title, the test is whether a reader could point at the exhibit and check it.

P1 (guided). Test three titles against one chart (India). A chart shows Turant's contribution per order across eight quarters: −38.21, −29.55, −19.48, −10.98, −8.41, −2.14, 4.57, 12.37 rupees. Fixed costs in the eighth quarter were ₹44.60 crore and order volume 1.62 crore. Three titles are proposed: (a) "Contribution margin by quarter"; (b) "Unit economics have turned and Turant is now profitable"; (c) "Contribution per order turned positive in Q7 and reached ₹12.37 in Q8". Score each against the three overstatement tests and name the winner.

Solution. Title (a) fails before the tests, because it is a dimension label with no verb and no claim, so there is nothing to test. Title (b) passes the recomputation test on its first half and fails the scope test on its second: the chart contains no fixed costs, and at ₹12.3680 per order against ₹44.60 crore of quarterly fixed cost, breakeven needs 44.60 / 12.3680 = 3.6061 crore orders, which is 3.6061 / 1.62 = 2.2260 times the current volume. The quarter's EBITDA is 20.0362 − 44.60 = −₹24.5638 crore. Title (c) passes all three: the numbers recompute, the scope is contribution per order and nothing wider, and the strength is a statement of fact about two labelled points. Title (c) wins, and (b) is the dangerous one because it is nearly right.

P2 (guided). Decide whether to split (US). Nine sector series over twelve months. Compute the number of pairs, and given that they cross 31 times, state the average crossings per pair. The dispersion between the highest and lowest index is 13.22 points at month 8 and 28.15 at month 12. State what share of the year's dispersion arrives after month 8 and say whether you would split the chart into panels.

Solution. Pairs are 9 × 8 / 2 = 36, so crossings average 31 / 36 = 0.8611 a pair. The dispersion created after month 8 is (28.15 − 13.22) / 28.15 = 53.03%, so more than half the year's separation happens in the last four months. Split it. Two reasons, and the second is the stronger. Thirty-one crossings make single-line tracking expensive for a reader. And because the series sit on top of each other for two thirds of the year, the crowded chart's most legible period is the part where the least is happening, which is the opposite of what the message needs. Nine panels on a shared scale, ordered by full-year return.

P3 (guided). Repair four label titles (India and US). Rewrite each as an action title using the figure given. (a) "Contribution margin by city", where six of the seven named cities are positive in the latest quarter and Chennai is at −₹9.75 against a network ₹12.37. (b) "Sector performance", where eight of nine sectors ended the year positive and health care ended at −1.85%. (c) "ARR and retention", where year-on-year growth fell 26.61 points while gross retention rose 6.7 points. (d) "Turnaround times", where the network breach rate is 5.71% against 4.12% the month before.

Solution. (a) "Six of the seven named cities now earn a positive contribution per order; Chennai does not, at −₹9.75 against a network ₹12.37." (b) "Eight of the nine sectors ended the year higher; health care alone ended lower, at −1.85%." (c) is a trap, because the two facts are two messages and no single title should carry both; the correct answer is to split, with "Year-on-year growth fell 26.6 points over four quarters" on one exhibit and "Gross retention rose 6.7 points, so the slowdown is in new business" on the next. (d) "Turnaround breaches rose to 5.71% of tests from 4.12%, above the 4.00% standard." Note what none of them says: "significantly", "sharply" or "materially", each of which is an adjective standing where a number should be.

P4. The sign-change trap (US and India). Contribution per order was −₹2.14 in the sixth quarter and ₹12.37 in the eighth. An analyst writes "margin tripled in two quarters". Compute the three candidate figures an analyst might have used, say why each fails, and give the statement you would put on the slide.

Solution. The ratio is 12.3680 / (−2.1350) = −5.7930, which is negative and cannot be a growth multiple. The percentage change against the absolute base is (12.3680 + 2.1350) / 2.1350 = 679.30%, an operation a spreadsheet performs happily and which means nothing, because a percentage change measures movement relative to a base that has to be on the same side of zero. The one-quarter ratio 12.3680 / 4.5700 = 2.7063 is arithmetically valid but is a different claim about a different pair of quarters, off a base of ₹4.57. The slide says: "Contribution per order improved ₹14.50 over two quarters, crossing zero between Q6 and Q7."

P5. Split a dual-axis chart (US). ARR runs $41.2m to $73.9m over eight quarters and gross retention runs 86.4% to 93.1%. Someone has put them on two axes. Name the two exhibits you would build instead, give each an action title, and state the one number that makes the second title defensible.

Solution. Exhibit one is a column chart of net ARR additions, 5.6, 6.1, 5.5, 4.7, 4.1, 3.4, 3.3 $m, titled "Net ARR additions peaked at $6.1m in Q3 and have fallen every quarter since, to $3.3m", which is a fall of 3.3 / 6.1 − 1 = −45.90% from the peak. Exhibit two is a line of gross retention on an axis from 85% to 95%, titled "Gross retention improved 6.7 points to 93.1%, so the deceleration is coming from new business rather than churn". The number that makes the second title defensible is the retention improvement itself. If retention rose while net additions fell, the shortfall must be in gross new business, and the exhibit that proves it, a gross-new-against-churn split, belongs in the appendix so the claim can be defended when challenged.

P6. Price free scaling (US and India). A nine-panel grid covers a shared range of 29.7320 index points. Health care's own range is 3.5742 points and utilities' is 4.9884. Compute the exaggeration factor for each under free scaling. Then do the same for a sparkline column with a shared span of ₹46.72 where Chennai's own range is ₹14.92.

Solution. Health care 29.7320 / 3.5742 = 8.32 times; utilities 29.7320 / 4.9884 = 5.96 times; Chennai 46.72 / 14.92 = 3.13 times. The consequence is the same in both cases and worse in the second: on a panel grid a determined reader could compare tick labels and discover the axes differ, while a sparkline has no axis labels at all, so free scaling there is undetectable by the reader and the exaggeration is invisible.

P7. Cut a wall to a dashboard (India). A monthly page carries fourteen tiles. Compute the pairwise comparisons a reader can make at fourteen tiles and at six, the reduction, and the scan cost at four seconds a tile. Then name what has to be true of each surviving tile.

Solution. 14 × 13 / 2 = 91 comparisons against 6 × 5 / 2 = 15, a reduction of 1 − 15/91 = 83.52%. At four seconds a tile the scan cost falls from 56 seconds to 24. Each surviving tile carries three things: the metric now, a comparison that makes it interpretable, such as the prior period, the target, or the same metric elsewhere, and a status that states the judgement, so that ten readers leave with the same view rather than ten. The tile budget is fixed and a new tile displaces a named old one, which is the only mechanism that stops the page growing back.

P8. Count the focal points (India). A draft chart carries: value labels on all eight points, a red zero line, a shaded band over the negative region, a second accent on the fifth point, a legend for one series, and a callout box with the breakeven multiple. Count the elements competing for first attention and the pairwise comparisons they force, then give the repair order and what survives.

Solution. Six elements compete, forcing 6 × 5 / 2 = 15 pairwise comparisons before the reader thinks about the data at all. Repair order, working from the message backwards: delete the callout box, because it makes a claim the chart cannot support and belongs on its own exhibit; delete the shaded band, which encodes nothing the zero line does not; delete six of the eight value labels, keeping the endpoint; remove the legend, since one series needs none; and take the colour off the zero line, making it a thin grey rule. What survives is the accent on the segment where the line crosses zero, one endpoint label and one short annotation. One element, zero comparisons.

P9. Write a caption that works (US). A scatter of fourteen industrials fits a line with slope 0.1179 and R-squared 0.6887. Meridian Fluid sits at 22.0% return and 1.10x invested capital against a fitted 2.5695x. Refitting without Meridian gives an R-squared of 0.9932. Write the caption and say which verb you rejected and why.

Solution. "Exhibit 4. Companies earning higher returns on capital trade at higher multiples of it; Meridian Fluid, at 22.0% return and 1.10x invested capital, trades 57.2% below the fitted relationship. Source: synthetic dataset, latest reported fiscal year." The discount is 1.10 / 2.5695 − 1 = −57.19%. The rejected verb is "drive": fourteen companies at one date establish an association, not a direction, and the design cannot support a causal claim whatever the fit. The R-squared rising from 0.6887 to 0.9932 on the refit is the fact that justifies the exhibit's page, because it says thirteen companies sit on a line and one does not, which a correlation reported in prose would have hidden.

P10. Find the exhibit that adds nothing (India). A five-title sequence reads: (1) "Contribution per order turned positive in Q7 and reached ₹12.37 in Q8"; (2) "Contribution margin has improved every quarter since Q1"; (3) "Quarterly EBITDA loss narrowed from ₹71.06 crore to ₹24.56 crore and has never been positive"; (4) "Breakeven needs 3.61 crore orders a quarter, 2.23 times the current 1.62 crore"; (5) "Cash of ₹186 crore covers 18.0 months at the current ₹10.32 crore monthly burn". Identify the weak title, say why, and give the replacement.

Solution. Title 2 adds nothing. A reader who has accepted title 1, which states a crossing and an endpoint, already knows the series improved, and the eight-quarter path is visible on exhibit 1 itself. Delete it or replace it with the question exhibit 1 actually provokes, which is whether the improvement is structural: "Three fifths of the ₹50.58 swing came from gross margin, two fifths from lower variable cost", with 31.28 + 19.30 = 50.58 as the completeness check. The general test is that each title must be unpredictable from the one above it, and a title beginning "and it has been improving" is almost always predictable.

P11. Page, appendix or cut (India). Eleven candidate exhibits exist, of which five move the funding decision, four answer questions a director would plausibly ask, and two do neither. Build times are 25, 35, 30, 20 and 15 minutes for the five; 40, 45, 35 and 50 for the four; and 20 and 30 for the two. Compute the build time in each group and the share saved by cutting, then state the two properties an appendix must have.

Solution. Pages 25 + 35 + 30 + 20 + 15 = 125 minutes; appendix 40 + 45 + 35 + 50 = 170; cut 20 + 30 = 50; total if all were built 345. Cutting the two saves 50 / 345 = 14.49% of the build. The appendix needs two properties to work. One is an index page at the front, so any exhibit is reachable in one move. The other is an action title on every appendix exhibit, because the moment one is displayed it is doing a page exhibit's job and a label title costs the presenter the twenty seconds the title should have carried.

P12. The crown arithmetic (US). A five-exhibit set scores 0.94, 0.92, 0.91, 0.95 and 0.88 on the machine, and the learner committed a rubric self-score of 0.84 before the reveal. Compute the composite, say whether it passes a crown gate, and state the two ways to close the gap with the numbers for each.

Solution. Machine mean is (0.94 + 0.92 + 0.91 + 0.95 + 0.88) / 5 = 0.92, so the composite is 0.6(0.92) + 0.4(0.84) = 0.888, which fails the 0.90 crown gate by 0.012. Route one is the machine score. At a rubric of 0.84 the machine mean has to reach (0.90 − 0.336) / 0.6 = 0.94, which means fixing the 0.88 exhibit and more besides, since replacing that one exhibit with a 0.94 lifts the mean to 0.932 and the composite only to 0.8952. Route two is the rubric. At a machine mean of 0.92 the rubric would have to be (0.90 − 0.552) / 0.4 = 0.87, and the rubric is a committed self-score, so raising it means really improving the judgements it measures. Rescoring after the reveal doesn't count. The fair reading is that the weak exhibit is a symptom and the set needs a pass on all five.

Applied mini-project

Five exhibits, one sequence, from an analysis you have already done. Budget about four hours, of which about ninety minutes is before anything is drawn. If you have no prior analysis, use the fallback brief after the steps.

Step 1: write the five claims (about 45 minutes). Before opening any tool, write five sentences on one page. Each is a claim, each has a verb, each carries the number that makes it checkable, and together they answer the question the analysis exists to answer. Then read them in order with nothing else visible and check three things: that each says something the previous one did not, that a reader who accepted the previous one is ready for this one, and that deleting any one leaves a hole. Rewrite until all three hold. Most people cut two claims and add one at this step, which is the cheapest cutting you will ever do.

Step 2: recompute every number in the five claims (about 30 minutes). Not check, recompute, from the source, by a route that does not reuse the intermediate you already have. Then run the scope test on each claim, naming the period, the entity and the unit, and confirm all three sit inside the evidence you will show. Then run the strength test, looking specifically for a causal verb on correlational data, a level-rate confusion, and any ratio or percentage change computed across a sign change. Record what you found, including the claims that survived unchanged.

Step 3: specify the five exhibits (about 15 minutes). For each claim, write one line naming the message type, the form that carries it, the axis units, and the single element that will hold the accent. Do this before drawing, because the specification is what stops the exhibit growing a second message while you build it. Where two claims want the same form on the same data, check whether you have one claim written twice.

Step 4: build, with one focal point each (about 90 minutes). Build the five in the Studio. As each is finished, count the elements competing for first attention and record the number, then strip until it is one. Every exhibit gets its claim as its title, at a precision you can defend, within two lines.

Step 5: score, decide the appendix, and submit (about 60 minutes). Commit a rubric self-score for each exhibit before revealing the machine score, then run the seven-rule check and record both. Compute the composite for the set at 0.6 machine plus 0.4 rubric. Then take every other exhibit your analysis produced and sort it: does it move the decision, or would somebody ask for it, or neither. Submit the five exhibits with their scores, the appendix list with an index, the cut list with reasons, and five lines of sequence rationale saying what each exhibit adds.

The fallback brief. Turant's board must decide whether to approve the next funding round now or in two quarters. You have the eight-quarter unit-economics table, the city table for the last six quarters, cash of ₹186 crore, and a quarterly cash burn of ₹30.96 crore. Build the five-exhibit sequence that answers the timing question, using no data outside those tables, and state in one line what you would need that you do not have.

Rubric. Twenty-four points, passing at 20 with no zero on any row, and the crown threshold requires a composite at or above 0.90 as well.

Any exhibit whose title makes a claim its own body cannot prove scores zero on row 1 regardless of how the exhibit looks, and any number that cannot be reproduced from the submitted working scores zero on row 2. The set is graded on whether a reader who saw only the five titles would reach your conclusion, not on how the exhibits look.

Reading & resources

Core: do these.

• Cole Nussbaumer Knaflic, Storytelling with Data (2015): the closest published treatment of the one-message habit and the action title, with the clearest worked before-and-after sequences available anywhere. Read the chapters on focusing attention and on the "so what" first. [Paid] [Beginner]
• Edward Tufte, The Visual Display of Quantitative Information (1983, 2nd ed. 2001): the origin of small multiples as a named idea and the source of the argument that comparison of shapes beats tracking of lines. Read the chapters on data density and multifunctioning elements. [Paid] [Intermediate]
• Edward Tufte, Beautiful Evidence (2006): where sparklines are introduced and argued for, with the reasoning about data-word graphics that sit inside text. [Paid] [Intermediate]
• Stephen Few, Information Dashboard Design (2nd ed. 2013): the standard treatment of the dashboard-against-wall distinction, screen real estate, and the reading order of a monitoring page. Its taxonomy of dashboard failures is still the best checklist in print. [Paid] [Intermediate]

Extension: worth the time.

• Gene Zelazny, Say It With Charts (4th ed. 2001): the original message-to-form mapping from a consulting practice, and the source of the habit of writing the chart's message as its title. [Paid] [Beginner]
• Alberto Cairo, How Charts Lie (2019): the best single treatment of the overstatement failures, including scale manipulation and the causal reading of correlational pictures. Read it alongside the three strength tests. [Paid] [Beginner]
• Dona Wong, The Wall Street Journal Guide to Information Graphics (2010): a compact reference for the conventions of financial exhibits, including precision, labelling and source lines. [Paid] [Beginner]
• Jonathan Schwabish, Better Data Visualizations (2021): a broad catalogue of forms with a clear treatment of small multiples and panel ordering. [Paid] [Intermediate]

Free.

• The Financial Times Visual Vocabulary (open, updated periodically): a one-page map from message type to chart form, useful as a desk reference beside the form grammar. [Free] [Beginner]
• The BBC and Reuters graphics style notes, published on their engineering and graphics blogs: practical conventions for annotation, direct labelling and mobile-first exhibit sizing. Availability and location change, so search rather than bookmarking (checked September 2026; confirm the current URLs before relying on them). [Free] [Beginner]
• Storytelling with Data blog, exercise archive: makeover exercises with published solutions, which is the cheapest source of reps outside this branch's own drill set. [Free] [Beginner]
• Tufte's own essay collections and the ET forum archives: uneven but occasionally the clearest available argument about a specific convention. [Free] [Advanced]

Inside the programme, you'll find the declutter craft assumed here in the previous chart-craft module, and the form grammar in the one before that. Titles across a whole deck belong to the storytelling branch's second node. The written-report treatment of exhibits and captions belongs to the analyst-communication module in the reporting phase.

Teach it back & journal

Feynman prompt. Take a colleague who builds decks and has never been told any of this, and explain why two charts of identical data can be worth different amounts, without using the words design, beautiful or professional. Use the arithmetic. Nine series on one chart cross 31 times across 36 pairs, and the same nine in a shared-scale grid cross zero times, so a fact inside the first picture only becomes visible in the second. Six accented elements on one chart force fifteen pairwise comparisons before the reader thinks about anything. Fourteen dashboard tiles admit 91 comparisons and six admit 15. Then make the harder point: none of those numbers is about charts. Each one is about how much work the author did before drawing, and each is a cost the reader pays for work the author didn't do. If your colleague can tell you, unprompted, why a percentage change across zero is meaningless, you've taught the idea and not only the arithmetic.

Journal prompt. Open the last deck or report you produced and pull out every exhibit. For each one, write its title as it appears, then write, in your own words, the single claim you intended it to make. Count how many of your titles are the claim and how many are a dimension label with the claim existing only in your head. Then take the three exhibits you were most pleased with and run the three overstatement tests on their titles: recompute every number, state the period, entity and unit each claim covers, and check for a causal verb, a level-rate confusion or a ratio across a sign change. Write down what you found, including the claims that survived. A habit only becomes a habit when you have evidence it catches things. Then answer one uncomfortable question truthfully. When you built those exhibits, did you write the sentence first or the chart first? Most people find they wrote the chart first for every one, and the reason isn't laziness. Writing the sentence first means committing to a claim before the picture is there to hide behind. Finish by choosing one exhibit from that deck and rebuilding it with the title written first. Then note what changed about the chart itself, because in most cases the chart that supports a stated claim isn't the chart you'd have drawn from the data alone.

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