Diagramming Ideas: Concept, Process, Flow, Frameworks, Matrices, Gestalt

Presentation track, branch pr-visual (Diagrams & Storyboards) · Node PR3.01 · ~10 focused hours · Mastery gate ≥ 85%

A diagram on a slide carried nine boxes, fourteen arrows, nine icons, three floating labels, a legend, nine drop shadows and a title. Forty-six objects. The analyst who drew it could talk anyone through it in ninety seconds, which is how she knew it worked. Nobody in the room could read it without her, which is how everyone else knew it did not.

She had drawn what she understood. A diagram has to draw what the reader must understand, and those are different pictures. Hers had an arrow for every relationship she had discovered during three weeks of work, because each one had cost her something to find. The reader needed five of those fourteen arrows and would have been better served if the other nine had never been drawn.

A diagram is a claim about relationships. A table is a claim about values, and a chart is a claim about quantities. When what the reader must hold in their head is which parts touch which, in what order, and with what consequence, a table forces them to rebuild the picture themselves and a diagram hands it over intact. When what they must hold is how large something is, the diagram is decoration and the table was right all along.

Here is the arithmetic that makes the case concrete. Four banks in a retail payment system exchange money in twelve directions on one day. Give a reader that as a bulleted list of twelve legs and ask which bank has to fund settlement, and they must add three receipts, add three payments and subtract, for each of four banks. That's twenty operations, done in their head, while you keep talking. Print the same twelve legs as a diagram with each bank's net position on its node and the answer is one lookup. The information content is identical. The reader's cost differs by a factor of twenty.

Diagrams also have a property that most slide furniture lacks, which is that they can be checked. Money that leaves one node arrives at another. Goods that leave a mill reach a depot, reach a customer, or are written off as shrinkage. Every flow diagram carries a conservation law, and a conservation law is a test. Sum the four net positions in that payment example and you must get zero. Get anything else and the drawing is wrong, whatever it looks like, and you have found the error before your audience did.

The reason a good diagram feels effortless is that it is recruiting machinery the reader is not aware of using. Things placed close together are read as one group before any conscious reading happens. Things enclosed by a line are read as belonging together even when nothing else says so. Those responses are fast, involuntary and older than the slide. You are either using them or fighting them, and a diagram that fights them costs the reader real work to overcome.

There is a limit worth stating this early. Diagrams are poor at magnitude. Nine boxes of similar size say nothing about whether one flow is a hundred times another, so a payment diagram needs its numbers printed on the arrows instead of encoded in their thickness. And a great many diagrams should not exist at all. A six-stage cycle drawn as a wheel with icons and a legend spends forty-five objects to carry seven facts. Written out, the same seven facts take three sentences and forty-four words, and the sentences are better because they can say which stage matters and why.

Charts of data belong to the chart-craft branch, which owns form choice, decluttering and the action title. What is owned here is everything that is not a chart: the concept picture, the process map, the flow of money or goods, the framework rendered as a visual, and the two-by-two that has quietly decided more corporate priorities than any model. The grammar is different because the failure modes are different, and both are teachable.

Learning objectives

By the end of this node you can:

1. Decide whether a relationship belongs in a diagram, a table, a chart or a sentence, and defend the choice against the strongest alternative.
2. Name the five diagram families, state the question each one answers, and pick the family from the message rather than from a template gallery.
3. Draw a process diagram with correct step, decision and loop notation, and add swimlanes only when the actor changes what the reader should do.
4. Build a flow diagram of money or goods that satisfies its conservation check, and use the check to find a transcription error before the diagram ships.
5. Construct a two-by-two whose axes are defensible claims, state the cut rule for each axis, and report what the matrix hides about the underlying ranking.
6. Render a published framework as a visual without turning it into decoration, attributing the author and the book in the sentence that names it.
7. Apply proximity, similarity, enclosure, continuity and closure to make a crowded diagram readable, and measure the proximity ratio you achieved.
8. Align and space diagram elements on a stated grid, computing box widths and gaps from the canvas rather than nudging them by eye.
9. Use icons, labels, arrow verbs and a single accent colour so that every mark on the diagram carries information, and check the accent's contrast ratio.
10. Recognise the diagram that should have been three sentences, measure its objects-per-fact cost, and delete it.

Prerequisites & connections

Builds on. Nothing from outside the Diagrams & Storyboards branch is required. The arithmetic is percentages, weighted averages, one standard deviation and one correlation. If you've already worked PR1.03, the object mechanics will be familiar and you'll draw faster. If you've worked PR2.02, the colour rule below will read as the same rule applied to a different object. Neither is a gate, and the branch is built so you can come in here.

Feeds forward. PR3.02 takes the diagrams you can now draw and puts them inside a ghost deck, where each exhibit is only sketched. PR4.01 and PR4.02 order the claims your diagrams support. PR5.01 to PR5.03 supply the archetypes where process maps and two-by-twos actually turn up, and the consulting deck template in particular assumes you can draw one without help. PR7.01 runs the whole sequence on one analysis.

Four neighbours own material that touches this one, and the boundaries are deliberate. PR2.01 owns chart-form choice for data. So when a section below says a quantity comparison belongs in a sorted bar, it names the rule and moves on without teaching it. PR2.02 owns decluttering and colour as data-visualisation craft, including the one-accent formula and the contrast check. Both get applied here to diagram objects instead of chart marks, and the underlying rule isn't rebuilt. PR2.03 owns the action title and the one-message test for a chart, and you'll see the same test applied here to a diagram. PR1.03 owns object mechanics: grouping, the selection pane, shape defaults, animation and export. When a section below tells you to align five boxes on a computed grid, it gives you the arithmetic and assumes you know which menu does the job, because the mechanics node owns that menu.

One more seam runs sideways. CN1.02 owns the issue tree and hypothesis-driven structuring as consulting practice, and an issue tree is a hierarchy diagram. Here you learn to draw one so it reads. There you learn to make it exhaustive. A beautifully drawn tree that overlaps at the second level is still wrong, and no drawing rule will catch that.

Core teaching content

1. When a diagram is the right form

Start with the question your reader is holding. If it's "how big", "how many" or "how has it moved", the answer is a chart or a table, and a diagram will only add distance. If it's "how does this work", "who touches this and in what order", "where does the money go" or "what has to be true for the recommendation to hold", the answer is structural. Structure is what a diagram encodes natively.

To tell the two apart, ask whether the reader has to reconstruct anything. Read a table of twelve payment legs and you know twelve facts and no relationships. The relationships are all still to be assembled, and you assemble them by holding rows in working memory while you scan for the matching row. That reconstruction is the cost, and a diagram has already paid it. The twelve legs among four banks are the twelve ordered pairs that four nodes admit. Once they're drawn as arrows between four circles, the question "who pays whom" needs no memory at all.

So before you pick a shape, ask what the reader would have to reconstruct if you didn't draw it.

The same test rules diagrams out. Say the message is that one channel grew 169.7 percent while another grew 167.2 percent. There's no structure in that claim, just two magnitudes and a comparison, and a sorted bar carries it. Draw the two channels as boxes with an arrow between them and you've invented a relationship the data doesn't contain. That's the particular harm a wrongly chosen diagram does. It's worse than the harm of a wrongly chosen chart, because a reader who sees an arrow assumes somebody meant it.

There's a third outcome that authors resist, which is that the right form is a sentence. Structure is only worth drawing when there's enough of it. Three steps in a fixed order with no branches and no actors is a sentence, and a sentence read at 220 words a minute is faster than any picture of the same three steps. The working rule: below about four elements with no branching, write it. Above that, or wherever a decision or a loop exists, draw it.

Two secondary reasons can justify a diagram even when a sentence would serve. One is durability. A reader coming back to a deck three weeks later can re-enter a picture at any point, but has to re-enter prose at the beginning. The other is that a diagram is a shared object in a discussion, because people point at diagrams. Neither reason survives if the diagram is unreadable, so neither one excuses the forty-six-object version.

2. The five families, and the question each one answers

Five families cover almost everything you'll need to draw as an analyst. Memorise them as questions, because the question is what you already have and the shape is what you're choosing.

The families aren't interchangeable. The commonest authoring error is picking the family from what the software offers instead of from the question. A gallery of preset graphics puts cycles, pyramids and funnels near the top of its list, and each of those is a hierarchy or a process wearing an unusual costume. A funnel is a process whose stages happen to shrink. A pyramid is a hierarchy drawn so the top level looks important. A cycle is a process with its last arrow bent back to the first, and it only tells the truth when the loop is real, which is less often than the shape gets used.

Two families deserve a warning up front. A concept diagram has no built-in check, so it's the family where nine boxes and fourteen arrows happen. Nothing in the form pushes back on an extra arrow. A matrix carries a built-in claim that authors rarely notice they're making, which is that the two axes are independent enough to be worth two dimensions. The two-by-two section below measures that claim on a real project list and finds it fails.

Mixed forms exist, and they're usually a mistake. A process diagram with a small hierarchy hanging off step three is two diagrams merged to save a page. The saving is nominal, because the reader now has to work out which grammar applies where. Split it. Page count is cheap, and comprehension isn't.

5. The two-by-two: two axes, two claims, four consequences

A two-by-two matrix is the most persuasive diagram in business and the least examined. It turns a list into a set of recommendations without appearing to argue. That's its power, and it's also its danger. Every one has three decisions buried inside it: what the axes measure, where the cut on each axis falls, and what the four quadrants are called. Each is a claim, and each should be defended on the page.

Work it on a real-looking list. A US distributor is choosing among nine operational projects, each with an estimated annual benefit in $000 and an effort in person-weeks: EDI onboarding for the top twenty suppliers at 142 and 14, automated three-way-match tolerance at 91 and 4, a dock scheduling app at 38 and 9, barcode receiving at 64 and 6, a supplier scorecard at 22 and 3, invoice OCR at 118 and 11, carrier consolidation at 205 and 22, pallet standardisation at 47 and 5, and exception dashboards at 29 and 2. Benefits total 756 and effort totals 76 person-weeks.

The cut rule is the first hidden decision. Split at the medians, benefit 64 and effort 6, and you get two projects in "do now", three in "big bet", three in "fill-in" and one in "drop". Split at the means, benefit 84.0000 and effort 8.4444, and barcode receiving moves out of "do now" into "fill-in", because 64 clears the median but not the mean. So one project of nine changes its recommendation on a choice of summary statistic that nobody in the room was told about. That should bother you. A matrix that doesn't print its cut rule asks the audience to accept a recommendation whose derivation they can't see, so print it. A note reading "cut at the median of each axis (benefit 64, effort 6)" costs one line.

The second hidden decision is whether the axes are independent, and here the matrix's own data usually tells on it. The correlation between benefit and effort across these nine projects is 0.9095, so 82.72 percent of the variation in one axis is already carried by the other. The nine points fall close to a line running from bottom left to top right. So the two off-diagonal quadrants are thinly populated by construction, and the picture is closer to a ranking drawn on a diagonal than to a real two-dimensional space. Say that out loud in the annotation, because the reader's instinct is that four quadrants mean four distinct kinds of project.

The third problem is what the matrix hides. Rank the same nine projects by benefit per person-week and the order is: automated three-way-match tolerance at 22.7500, exception dashboards at 14.5000, invoice OCR at 10.7273, barcode receiving at 10.6667, EDI onboarding at 10.1429, pallet standardisation at 9.4000, carrier consolidation at 9.3182, the supplier scorecard at 7.3333 and the dock scheduling app at 4.2222. Exception dashboards are stuck in the unglamorous "fill-in" quadrant, yet they rank second of nine on the ratio. The top four projects by ratio deliver 302 of benefit for 23 person-weeks, or 13.1304 per week, against the "big bet" quadrant's 465 for 47 weeks, or 9.8936 per week. A quadrant is a bucket, and a bucket destroys the order inside it.

None of that makes the matrix useless. It makes it a way to communicate an analysis you've already done. So run the ranking first, decide from the ranking, and then draw the matrix to show a room why the decision groups the way it does. Add the ratio as a label on each point and you've kept the picture's speed and given back the ordering it threw away. If you can only have one, keep the ranking. If you can have both, the matrix is where the ranking gets agreed.

The quadrant names are the last claim, and they do the most work. "Do now, big bet, fill-in, drop" is a set of instructions, and a reader who accepts the axes has accepted the instructions too. Names like "high priority" and "low priority" pretend to be descriptive. They aren't. Choose names you'd defend in a sentence like "we are dropping the dock scheduling app because it is low benefit and high effort". If that sentence sounds wrong, either the quadrant is misnamed or the project is misplaced.

6. Frameworks as visuals, without turning them into decoration

Analysts inherit frameworks. Barbara Minto's pyramid, from The Pyramid Principle, is a hierarchy. Michael Porter's five forces, from Competitive Strategy (1980), is a concept diagram with a centre and five surrounding pressures. The growth-share matrix that the Boston Consulting Group popularised in the early 1970s is a two-by-two with named quadrants. Each arrives with a canonical picture, and that picture is where the trouble starts. Reproducing it is easy. Reproducing it usefully isn't.

A framework on a slide has to do one of two jobs, and it has to declare which. Either it's teaching the framework, in which case the generic picture with generic labels is correct and the slide is a lesson. Or it's applying the framework to the case in front of you, and then the generic labels are the failure. A five-forces diagram whose five boxes read "supplier power, buyer power, threat of entry, substitutes, rivalry" has told the reader nothing they couldn't have looked up. The same diagram with "three refiners control 71 percent of feedstock", "top four customers are 58 percent of revenue" and the other three boxes filled with the case's own facts is an analysis. The empty version is the commonest framework failure in corporate decks, and it's entirely a labelling failure. The geometry was fine.

The second failure is completeness for its own sake. A framework has five boxes, so five boxes get filled, including the two where nothing interesting is happening. Filling a box with "moderate" is worse than leaving it visibly thin, because "moderate" claims an assessment you didn't make. Draw the two forces that decide the answer at full size, draw the other three small with one line each, and let the asymmetry carry the finding. You don't owe the reader the framework's own shape.

Attribution costs one clause and settles a question the reader might otherwise ask. Name the person and the book in the sentence where the framework first appears, the way the paragraph above does. Don't reproduce a source's enumerated list as your own list, and don't quote more than a phrase. If you've modified the framework, say so in the caption: "adapted, with the two forces we found immaterial dropped". Used without attribution, a framework reads as either ignorance or theft, and six words avoid both.

There's one more move worth learning, and it's refusing the framework. Frameworks are lenses, and a lens that doesn't fit the case distorts it. If your five-forces analysis keeps producing boxes that say "not applicable", the industry may not be structured the way the framework assumes. Then the accurate slide says so and uses a different structure. Reaching for that refusal is a senior habit, and you can learn it. Run the framework in your notes and count how many boxes produced a finding. If it's fewer than half, don't put the framework on the page.

8. Alignment and spacing: the grid still rules

Diagram elements are placed either by arithmetic or by hand, and you can see the difference from the back of a room. Do the arithmetic once and you can reuse it forever.

Take a widescreen canvas of 960 by 540 points with a 48-point margin on each side. That leaves 864 points of content width. A twelve-column grid with a 24-point gutter gives a column width of (864 − 11 × 24)/12 = 50.0000 points and a column pitch of 74.0000 points. The check that closes it is 12 × 50 + 11 × 24 = 864.0000. Every box you'll ever place on that canvas is some whole number of columns plus the gutters between them, so define the grid before you draw anything.

Five equal boxes with a 24-point gap between them are (864 − 4 × 24)/5 = 153.6000 points wide. Their left edges fall at 48.0, 225.6, 403.2, 580.8 and 758.4. Their right edges fall at 201.6, 379.2, 556.8, 734.4 and 912.0, and that last right edge meets the content boundary at 912.0000 with a residual of zero. A box 90 points tall, centred vertically, has its top at (540 − 90)/2 = 225.0000. Those numbers take ten seconds on a calculator, and they're the difference between a row that looks built and a row that looks assembled.

Now compare a hand-placed version of the same five boxes, with lefts at 51, 228, 402, 585 and 761 and widths of 150, 156, 152, 149 and 155. The gaps between boxes come out at 27, 18, 31 and 27 points: a mean of 25.7500, a range of 13 and a sample standard deviation of 5.5000. Tops are at 214, 210, 216, 211 and 215, a range of 6 points with a standard deviation of 2.5884. Widths vary by 7 points with a standard deviation of 3.0496. The worst single left-edge error against the computed layout is 4.2000 points. No individual error is large. The effect is cumulative, though. Your reader's eye is registering three distinct gap values, five distinct widths and five distinct tops, which is thirteen small decisions to absorb before any content is read. The computed version presents one gap, one width and one top. That's three.

That count is the useful way to think about alignment, and it reaches well past diagrams. Every distinct value in a layout is something the eye has to judge as meaningful or not, and the judgement costs something even when the answer is "not meaningful". Cutting thirteen values to three takes ten judgements out of the reader's path to your content.

Two practical corollaries. When five boxes won't fit legibly, don't shrink the type. Compute the two-row variant. A top row of three on the same grid gives (864 − 2 × 24)/3 = 272.0000 points a box, nearly double the width, and two rows of legible boxes beat one row of unreadable ones every time. And when a diagram has to align with something already on the slide, such as a chart's plot area, align to that object's edges instead of the canvas. The reader sees the relationship between the two objects and not the relationship of either to a margin they can't see.

9. Iconography: signal against decoration

An icon is worth including when it identifies something faster than a word would, and in a business diagram that happens less often than icon libraries imply. A truck next to "carrier" is redundant. The word is already there, and it's read faster than the picture is recognised. A truck in place of the word "carrier" is different, provided the same truck appears on every carrier node and only on carrier nodes. It's a legitimate encoding. The reader learns the mapping once and then reads position instead of text.

That's the consistency test, and it has two halves. Every instance of a category gets the icon, and no instance of another category gets it. Half-applied iconography is worse than none. A reader who sees a truck on two of four carrier nodes concludes those two differ from the other two, then spends attention finding out they don't. The concept-diagram repair above deleted all nine icons for this reason. They were decorative, none of them encoded a category, and each one was costing an inspection.

Three more rules stop icons degrading a diagram. Use one visual family throughout, meaning one stroke weight and one level of detail, because a detailed illustration next to a flat glyph reads as two different kinds of object. Match icons by optical size instead of bounding box, since a wide glyph and a tall glyph at identical box sizes look different. And keep them monochrome in the diagram's neutral colour unless the icon is the accent, which is a decision the colour section below governs.

Licensing is a real question, and it belongs with the mechanics of shipping a deck. One line of it matters at drawing time, though. Check the licence before your diagram comes to depend on the icon. If you find out at export that a set can't be used commercially, you'll be redrawing instead of substituting. Icon-set licences change, so verify against the provider's current terms at the time you use them, and don't rely on what was true on a previous project.

10. Labels, and the arrow that needs a verb

Every mark in a diagram should be readable without the author present, and labels are where that promise is kept or broken. Three decisions do most of the work: where the label sits, what the arrow says, and what the diagram is called.

Put a label inside its shape when the shape is large enough to hold it at full type size, and outside when it isn't. Shrinking type to fit a box is the wrong trade, because you chose the box, and type size is a legibility floor you didn't choose. When labels have to sit outside, place them consistently, all above or all to the right, so the reader learns the convention once. A diagram where three labels sit above their nodes and two sit below has invented a distinction it doesn't mean.

Arrows are the most abused mark in business diagrams, because an unlabelled arrow seems to say something while committing to nothing. It might mean "sends money to", "reports to", "causes", "precedes", "influences" or "is related to", and those are six different claims with six different burdens of proof. Give every arrow a verb. "Pushes ₹42.80 crore to" is a claim. "Approves" is a claim. A bare arrowhead invites the reader to supply whatever relationship they already believed, so your diagram ends up confirming their prior instead of changing it.

Direction deserves the same care. If an arrow is bidirectional, say what flows each way. A double arrowhead usually hides two different relationships travelling in opposite directions, and drawing them as one mark loses both. In the four-bank payment diagram every pair exchanges value in both directions, so the correct treatment prints two numbers instead of one arrow with two heads.

The title is the last label, and it's the one that carries the argument. A title that says "process overview" is a filing label. A title that says "126 of 1,000 invoices leave the straight-through path and cost $118,800 a year" is the claim the diagram supports, and a reader who reads only titles still gets the finding. It's the same rule the chart-craft branch applies to charts, and it carries over unchanged. The title states the message and the diagram is the evidence. If you can't write the title, that's a sign the diagram has no single message yet.

Small print has one home, which is a footnote line under the diagram carrying the source, the period, the units and any exclusion. Put those on the diagram itself and they compete with the content. Put them nowhere and the diagram is unusable six weeks later, when someone asks which quarter it covers.

11. Colour in a diagram: one accent, measured

The colour rule from the chart-craft branch transfers without modification. One accent carries the message. Everything else recedes to a neutral. Colour that doesn't encode information is noise, and noise is expensive in a diagram, because a diagram has more objects than a chart does.

Measure the accent two ways, because the count and the area can disagree. Picture a process diagram with 13 shapes, five large boxes of 153.6 by 90 points, five medium of 120 by 72 and three small of 100 by 60. Its total shape area is 130,320 square points. Accent one large box and that's 1 of 13 shapes, or 7.6923 percent by count, and 13,824 of 130,320 square points, or 10.6077 percent by area. Both figures are in the range where an accent reads as an accent. The version this replaced used seven hues across those 13 shapes. That puts the most any single hue can claim at 14.2857 percent of area, and it marks nothing at all. When every object is coloured, colour has stopped being a signal and become a property of the diagram.

Contrast is the second measurement, and it's the one that fails without warning on a projector. The WCAG 2.2 ratio for normal-size text is 4.5 to 1 (checked September 2026; confirm, since guideline versions advance and the thresholds can move). Against it, a burnt-orange accent of #B34700 on a white ground gives 5.5002 to 1. White text on that same accent gives the same 5.5002 to 1, so the accent works as a fill and as a background. A neutral outline of #6B7280 on white gives 4.8345 to 1, and a dark label of #1F2937 gives 14.6791 to 1. The grey that authors reach for when they want something to recede, #9CA3AF, gives 2.5388 to 1 and fails. So the parts of the diagram you meant to be quiet are illegible to some of the room.

One pairing in that set deserves its own attention. The accent against the neutral, #B34700 on #6B7280, gives 1.1377 to 1. Those two colours are nearly identical in luminance and differ almost entirely in hue. A viewer with a common form of colour vision deficiency, or anyone looking at a projector with a washed-out gamut, may not tell them apart at all. So colour must never be the only encoding. Give the accented box the heavier outline as well, or the bold label, or the position at the start of the reading path. Encode twice and the diagram survives a bad projector, a monochrome print and a colour-blind reader, and you'll meet all three.

Two smaller decisions round this out. Diagram fills usually work better as very light tints of the neutral than as saturated colour, because a fill's job is to separate a box from the ground, and attracting attention isn't part of that job. And a diagram in a deck should use the same accent as that deck's charts. A reader who learned on slide 4 that orange means "the thing we are recommending" will apply that learning on slide 9, and rewarding the transfer costs you nothing.

Colour is the cheapest thing to add to a diagram and the most expensive thing for a reader to ignore.

12. The diagram that should have been three sentences

Some diagrams shouldn't be drawn at all, and the way to know is to count objects against facts.

A software vendor's six-stage buying cycle runs 21 days of awareness, 34 of research, 18 of shortlisting, 45 of evaluation, 29 of approval and 26 of onboarding, 173 days in total. Evaluation is the longest stage at 26.0116 percent of the cycle, and evaluation plus approval together are 74 days, or 42.7746 percent. Seven facts are in play, six stage durations and the total.

The version that arrived was a circular process wheel with icons. It carried 6 stage wedges, 6 stage labels, 18 sub-labels, 6 icons, 6 circular arrows, a centre ring, an outer ring and a legend. In all, 45 objects for 7 facts, which is 6.4286 objects a fact. Redraw it as a plain row of six boxes with durations printed on them and five arrows between, and you get 6 boxes, 6 labels, 6 duration labels and 5 arrows, which comes to 23 objects, or 3.2857 a fact. The wheel spends 22 more objects than the row, 95.6522 percent more, for identical information and one false claim. The cycle isn't a cycle. A customer who onboards doesn't go back to awareness.

Now write it out. "The buying cycle runs 173 days from first awareness to onboarding. Evaluation and approval together take 74 of those days, 42.8 percent of the total, and they are where deals stall. Everything before evaluation takes 73 days combined and is largely outside our control." Three sentences, 44 words. They carry the two facts that matter, plus an interpretation no drawing of six boxes can express. The row of boxes tells the truth cheaply. The sentences tell it more cheaply still, and they say which stage the reader should care about.

One rule falls out of this, and it's worth carrying. When a diagram's structure is a straight line and the interesting content is a comparison of magnitudes, the diagram is the wrong form twice over. A straight-line structure needs no picture, and magnitude comparison belongs in a sorted bar or in a sentence with the numbers in it. A wheel, a funnel or a pyramid drawn over a straight sequence adds a claim about the shape of the process that the data doesn't support. That added claim is the cost authors miss, because it doesn't show up as clutter.

Use the objects-per-fact ratio as a rough diagnostic. Somewhere above about four objects a fact, ask what the extra objects are buying, and be ready for the answer to be that the diagram should be a table, a bar chart or a paragraph. The three-sentence version of anything is always available, and it's almost never the version that gets built. That's the bias worth correcting.

13. Drawing the idea rather than the org chart

The last habit is the hardest to teach, and it's the one that separates a competent diagram from a useful one. Most business diagrams draw the structure that exists: the departments, the systems, the legal entities, the steps as the process document lists them. Very few draw the idea the reader needs to understand, and the two overlap less than you'd expect.

The receiving-to-pay case makes the difference concrete. The org chart version draws supplier, carrier, dock and accounts payable in four lanes with every step in each lane. It's accurate, and it answers "how does this work". The idea version draws one thing, the match decision, with 87.4 percent flowing one way in four steps and 12.6 percent flowing the other in nine, and the annual cost of $118,800 printed on the exception branch. The first is documentation. The second is an argument for changing a tolerance setting, and it fits in a quarter of the space.

To get the second version, ask what the reader must understand for the recommendation to land. That's a narrower question than what's true or what's complete. Every element the claim doesn't rest on is competing for attention with the elements it does rest on, and the elimination test is direct: cover an element with your thumb, and if the claim still holds, the element was decoration.

Completeness is a property of your understanding. Relevance is a property of the reader's question, and only one of the two belongs on the slide.

That gives you a two-version habit worth building on purpose. Draw the complete picture first, because you need it to be sure you understand the system. Then draw the argument version from it, keeping only what carries the claim. The complete version goes in the appendix, where questions can be answered from it. The argument version goes on the slide. Authors who skip the first version draw arguments that turn out to be wrong under questioning. Authors who ship the first version present documentation and wonder why nothing happens.

There's one situation where the org chart is the idea, and it's worth naming so the rule doesn't turn into a superstition. Sometimes the finding is about the structure itself, such as a process that crosses four organisational boundaries and fails at three of them. Then the structure is the argument, and drawing it fully is correct. The test still applies. You're drawing the structure because the claim is about the structure, not because the structure is what you happen to have.

Half the elements makes a useful forcing constraint while you're learning this. Take a finished diagram, rebuild it at half the object count, and put the two side by side. Usually the smaller one is better. When it isn't, you'll be able to say which element you missed and why the claim needed it, and that's the knowledge the exercise exists to produce.

14. Where the assistants help, and where they are dangerous

Current chat assistants are useful at three narrow tasks here, and dangerous at a fourth. They're good at listing the relationships implied by a description, which turns up the ones you left out. They're good at writing a text description of a diagram's structure that you can check for contradictions. And they're reasonable at proposing the family from a clearly stated message, though they'll over-recommend the process family because it's the most represented in their training material.

The dangerous case is the one that looks most helpful, which is asking for the diagram itself. An assistant will produce a plausible structure with plausible arrows and plausible labels, and the arrows will assert relationships nobody verified. A diagram's arrows are claims, and a claim generated just to complete a picture is the very defect you're learning to prevent. Use the assistant to attack a diagram you drew, not to draw one.

{
  "skill": "Stress-testing a diagram's structure and its conservation check before it is drawn or shipped",
  "use": "After you have drawn the diagram yourself, describe it to the assistant in text (nodes, arrows with their verbs and values, groups, title) and ask it to attack the structure: find arrows that assert an unverified relationship, find the conservation identity and test it, and name the elements that could be deleted without weakening the title's claim. Never ask it to generate the diagram.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "A spreadsheet or Python for the conservation arithmetic", "Your presentation tool, opened only after the structure survives"],
  "prompt": "Here is a diagram I drew, in text. Title: <TITLE>. Nodes: <LIST WITH ANY VALUES>. Arrows: <FROM, TO, VERB, VALUE for each>. Groups: <ENCLOSURES>. Do five things. 1) For each arrow, state the claim it makes in one sentence and mark any whose relationship my description does not actually establish. 2) State the conservation identity this diagram should satisfy, if any, and compute the residual from my numbers. 3) Count the objects and the distinct facts carried, and give objects per fact. 4) Name every element that could be deleted without weakening the title's claim. 5) Say whether a table, a sorted bar or three sentences would carry this better, and why. Do not draw or redesign the diagram.",
  "verify": "Recompute the conservation residual yourself from the source data rather than accepting the assistant's arithmetic, and require it to be exactly zero or to equal a named, printed reconciling item; an assistant that reports a clean residual on numbers it retyped has verified nothing. Check every arrow the assistant leaves standing against your own source: the test is that you can name the document, table or interview that establishes the relationship, and an arrow you cannot source comes off the page. Where it proposes deleting an element, delete it and read the diagram again yourself, because the deletion that a text description survives can still break a visual reading path.",
  "diy": "You must be able to choose the family from the message, draw the process notation correctly, compute the grid arithmetic and state the conservation identity with no tool open, because those are the four skills the mastery gate tests and the four an assistant will happily fake. The arrow-level judgement is the one that will be questioned in a room, and defending an arrow you did not decide to draw is not possible. An assistant that generates your diagram has authored the claims you will be asked to stand behind.",
  "market": "US"
}

Common mistakes & how experts think differently

1. Drawing every relationship you found. Three weeks of work produces fourteen relationships among nine nodes, and each one cost something to discover, so deleting one feels like a loss. Nine nodes admit 36 pairs, and 14 arrows hand the reader 14 things to inspect. The expert draws the five that carry the claim and keeps the rest in a note for the questions.

1. Hiding a decision inside a step. "Match invoice to PO and receipt" reads as one rectangle and is really a rectangle plus a diamond. Novices draw the happy path, because that's how the process is described to them. Experts ask what percentage takes each branch, and they discover that the 12.6 percent path holds most of the cost. If your process diagram has no diamond, either the process has no branches, which is rare, or you haven't found them.

1. Shipping a flow diagram with no conservation check. A quantity that moves has to arrive, be consumed, or be lost. Sum the nets and print the residual. A diagram whose nets sum to −1.30 instead of zero is wrong, and no amount of looking at it will reveal that. That's what the arithmetic is for.

1. Letting the cut rule decide the recommendation without saying so. Median and mean splits on the same nine projects move one project between quadrants, and nobody in the room is told which rule produced the picture. The expert prints the cut on the axis and also prints the ranking the quadrants discard, because a bucket destroys the order inside it.

1. Treating the two-by-two axes as independent when they aren't. A benefit-effort correlation of 0.9095 means 82.72 percent of one axis is already carried by the other, and two quadrants are thin by construction. Experts check the correlation before drawing four quadrants. When it's high they draw the ranking and use the matrix only to show a room why the ranking groups the way it does.

1. Reproducing a framework with generic labels. Five boxes reading "supplier power, buyer power, threat of entry, substitutes, rivalry" teach the framework and analyse nothing. The reader could look that up. Fill each box with the case's own facts, and draw the two that decide the answer larger than the three that don't. Then name the author and the book in the sentence that introduces the framework.

1. Nudging objects into place by eye. Five hand-placed boxes produced gaps of 27, 18, 31 and 27 points and widths spread over 7 points. That gave the eye thirteen distinct values to absorb, against three in the computed version, even though no single error exceeded 4.2 points. The expert computes the box width from the canvas. It takes ten seconds and never drifts.

1. Using colour as the only encoding. An accent and a neutral that differ in hue but not in luminance measure 1.1377 to 1, which is no difference at all to a colour-blind reader or a washed-out projector. Encode twice, so the accent also gets the heavier outline or the leading position. The habit costs nothing, and it removes a whole class of failure you'll never be there to see.

1. Drawing what exists instead of what matters. The complete four-lane process map is accurate and answers "how does this work". The one-node version with 87.4 percent, 12.6 percent and $118,800 answers "what should we change". Experts draw the complete version to be sure they understand it, then draw the argument version from it and put the complete one in the appendix.

1. Reaching for a wheel, funnel or pyramid over a straight sequence. Those shapes assert something about the process, that it cycles, narrows or ranks, and a six-stage buying cycle that ends in onboarding doesn't cycle. The wheel version spent 45 objects on 7 facts against the row's 23. The three-sentence version carried the same facts, plus the interpretation neither drawing could express.

Worked examples

Seven examples, three set in India and four in the United States. Every figure printed below comes from the computation restated at full precision in the Verify block at the end of the section. All the companies and banks named are synthetic. The data is constructed to be internally consistent and isn't drawn from any institution's published figures.

Worked example 1: A four-party payment flow, conserved and checked (India, ₹)

Four banks in a retail payment system settle one day's customer-to-customer transfers. The message you have to carry is which bank has to bring cash to the settlement window, and how much. All twelve ordered pairs carry traffic, in ₹ crore. Meru sends 42.80 to Sarayu, 18.60 to Nila and 9.40 to Tamra. Sarayu sends 31.20 to Meru, 26.50 to Nila and 14.30 to Tamra. Nila sends 22.90 to Meru, 19.60 to Sarayu and 11.80 to Tamra. Tamra sends 16.40 to Meru, 12.70 to Sarayu and 20.30 to Nila. The day carried 420,000 transactions.

Solution. Step 1, compute the gross positions. Sum each bank's outgoing legs and each bank's incoming legs. Meru sends 70.80 and receives 70.50. Sarayu sends 72.00 and receives 75.10. Nila sends 54.30 and receives 65.40. Tamra sends 49.40 and receives 35.50. Gross value is ₹246.50 crore measured either way, and that's the first check. The total sent must equal the total received, because every payment has a single recipient.

Step 2, net each node. Net equals received minus sent: Meru −0.30, Sarayu +3.10, Nila +11.10, Tamra −13.90.

Step 3, run the conservation check. The four nets sum to 0.0000000000. If they didn't, a leg has been transcribed wrongly, and this is the only place the error will announce itself. To see it work, misread Nila's net as +9.80 instead of +11.10, which is what dropping ₹1.30 crore from one incoming leg would do. The drawing looks unchanged. The nets now sum to −1.30, and the diagram is provably wrong before anyone has looked at it closely.

Step 4, derive the message. Two banks owe on net, Meru and Tamra, totalling ₹14.20 crore. Sarayu and Nila are owed the same ₹14.20 crore between them. So settlement funding is ₹14.20 crore against ₹246.50 crore of gross payments, and netting efficiency is 1 − 14.20/246.50 = 94.2394 percent. The average ticket is ₹246.50 crore over 420,000 transactions, or ₹5,869.05.

The wrong version. The same content as a bulleted list of twelve legs, one line each, under the title "Interbank payment flows". Every fact is present and nothing is false. A reader who wants the funder has to add three receipts, add three payments and subtract, for each of four banks. That's five operations a bank and twenty in total, done in their head while the presenter keeps going. Most readers don't attempt it, so the slide gets nodded at and the finding is lost. The list has no error check either. The transcription slip that changes Nila's net is invisible in a list, because a list of legs never asks the legs to sum to anything.

The corrected version. Four circles in a square arrangement, each labelled with the bank's name and its net position, with the net printed larger than the name and the two funders outlined in the accent colour. Twelve arrows, each labelled with its value, arrowheads carrying direction. A footnote line reading "Net positions sum to 0.00. Settlement funding ₹14.20 crore on ₹246.50 crore gross, 94.2 percent netting efficiency. Synthetic data, one settlement day." The title is the message: "Tamra funds 98 percent of the day's settlement, ₹13.90 crore of ₹14.20 crore."

Read it. The 94.2394 percent netting efficiency is the number the diagram exists to deliver. It's a statement about this one day and this one set of four banks, and it tells you nothing about the payment system in general. Nor does anything here say whether ₹14.20 crore is large for Tamra. That's the next question, and it needs a different exhibit. The conservation check proves the arithmetic is internally consistent. It doesn't prove the underlying legs were captured correctly, since two compensating transcription errors would still sum to zero.

Run this checklist on any flow diagram you draw. State the conservation identity and print the residual. Put the derived quantity on the node in place of the raw one. Label every arrow with a value and a direction. Give the title the finding instead of the topic, and add one line of scale context, such as the average ticket, to retire the obvious question.

Worked example 2: A receiving-to-pay process, with the decision node priced (US, $)

Kestrel Supply, a synthetic mid-size distributor, processes 1,000 supplier invoices a month across four parties: the supplier, the carrier, the receiving dock and accounts payable. The three-way match of invoice to purchase order to receipt succeeds 87.4 percent of the time. A straight-through invoice takes 4 steps and 3.0 days. An exception takes 9 steps and 11.5 days. Exception handling costs $18.40 an invoice. The average invoice is $4,860.00 on 2/10 net 30 terms, and 78 exceptions a month close after day 10. The message you have to carry is that the exception branch, and not the length of the process, is where the money is.

Solution. Step 1, split the volume. 874 invoices flow straight through and 126.0 become exceptions.

Step 2, compute the weighted averages that a summary would report. Steps: 0.874 × 4 + 0.126 × 9 = 4.6300. Cycle days: 0.874 × 3.0 + 0.126 × 11.5 = 4.0710. Both look healthy, and both conceal the distribution entirely.

Step 3, price the exception branch. Handling is 126.0 × $18.40 = $2,318.40 a month, or $27,820.80 a year. The discount at stake is 2 percent of $4,860.00, which is $97.20 an invoice. So 78 late closures cost 78 × $97.20 = $7,581.60 a month, or $90,979.20 a year. The discount leak is 7,581.60/2,318.40 = 3.2702 times the handling leak. That's the finding, because every instinct in the room will be to attack handling cost. The total leak at that one node is $118,800.00 a year.

Step 4, size the fix. Raising the match rate to 94.1 percent by widening the price and quantity tolerance takes exceptions from 126.0 to 59.0, a fall of 67.0. Handling falls to $1,085.60 a month, and missed discounts scale with volume to 37, or $3,596.40 a month. Annual saving is $62,616.00.

The wrong version. A single left-to-right chain of thirteen rectangles titled "Procure to pay process", with no diamond and no lanes, and the exception steps drawn inline as if everyone went through them. It's the version the process documentation produces. It says the process has thirteen steps, which is true of nobody: 87.4 percent of invoices take four and 12.6 percent take nine. It contains no branch, so it can't show a branch probability, so it can't carry a cost. The reader leaves with an impression of length and no finding about tolerance settings.

The corrected version. Four horizontal lanes, ordered by first appearance: supplier, carrier, receiving dock, accounts payable. Five handoffs, each crossing a lane boundary. One diamond, in the accounts payable lane, labelled "three-way match". Two exits, the upper one labelled "87.4 percent, 4 steps, 3.0 days" running to "pay on terms", the lower one labelled "12.6 percent, 9 steps, 11.5 days" running through the exception loop and back to the match step. The exception branch is the only accented element on the page and carries the annotation "$118,800 a year: $27,821 handling, $90,979 lost discounts". Title: "One in eight invoices misses the match and costs $118,800 a year, three-quarters of it in lost early-payment discounts."

Read it. The 3.2702 ratio is what turns this from a process map into a recommendation. It shows that the expensive part of an exception is the delay, and the work is the smaller cost. That conclusion depends on the 2/10 net 30 terms holding across the supplier base, so state that assumption on the slide. On net 30 with no discount, the discount leak is zero and the whole finding collapses to a $27,821 handling problem. The $62,616 saving also assumes the wider tolerance doesn't admit errors that cost more than the discount saved. It's the control question the finance team will ask, and this diagram deliberately doesn't answer it.

For a process diagram, the checklist runs like this. Every branch is a diamond with labelled exits, and every exit carries its share. Lanes are ordered by first appearance and exist only because handoffs matter. Loops return to the specific step that repeats. The costly branch is the accented one, and the title names that branch and its cost.

Worked example 3: The same project list as a ranking and as a two-by-two (US, $000)

Kestrel Supply must choose among nine operational projects for the coming half-year. Benefits are annual, in $000, and effort is in person-weeks. EDI onboarding for the top twenty suppliers, 142 and 14. Automated three-way-match tolerance, 91 and 4. Dock scheduling app, 38 and 9. Barcode receiving, 64 and 6. Supplier scorecard, 22 and 3. Invoice OCR, 118 and 11. Carrier consolidation, 205 and 22. Pallet standardisation, 47 and 5. Exception dashboards, 29 and 2. The message you must carry is what to start in the next six weeks.

Solution. Step 1, compute both possible cut rules before choosing one. Benefits sorted are 22, 29, 38, 47, 64, 91, 118, 142, 205, giving a median of 64 and a mean of 84.0000. Effort sorted is 2, 3, 4, 5, 6, 9, 11, 14, 22, a median of 6 and a mean of 8.4444. Totals are 756 of benefit for 76 person-weeks.

Step 2, place the projects under each rule and compare. On the median split, automated tolerance and barcode receiving go in "do now". EDI onboarding, invoice OCR and carrier consolidation go in "big bet". Supplier scorecard, pallet standardisation and exception dashboards go in "fill-in", and the dock scheduling app goes in "drop". On the mean split, barcode receiving moves from "do now" to "fill-in", because its benefit of 64 clears the median of 64 but not the mean of 84.0000. So one project of nine changes its recommendation on a choice nobody was told about.

Step 3, test whether the two axes deserve two dimensions. The correlation between benefit and effort is 0.9095, so 82.72 percent of the variation in one is already present in the other. The points lie close to a line, the off-diagonal quadrants are thin by construction, and the picture is a ranking on a diagonal wearing four quadrant names.

Step 4, produce the ranking the matrix hides. Benefit per person-week: automated tolerance 22.7500, exception dashboards 14.5000, invoice OCR 10.7273, barcode receiving 10.6667, EDI onboarding 10.1429, pallet standardisation 9.4000, carrier consolidation 9.3182, supplier scorecard 7.3333, dock scheduling 4.2222. The top four by ratio deliver 302 of benefit for 23 person-weeks, 13.1304 a week, against the whole "big bet" quadrant's 465 for 47 weeks, 9.8936 a week. Exception dashboards rank second of nine, and they're filed under "fill-in".

The wrong version. A two-by-two with unlabelled axes reading "impact" and "effort", no cut values printed, nine dots of equal size, quadrants named "quick wins, strategic, low priority, avoid", and the title "Project prioritisation". It recommends three big bets and two quick wins, and nobody can audit it, because a reader who disagrees has nothing to disagree with. Change the cut rule and the recommendation changes. Nothing on the slide reveals that a rule was chosen.

The corrected version. The same nine points, with the axes labelled "annual benefit, $000" and "effort, person-weeks", the cut lines drawn at the medians and annotated "cut at median: benefit 64, effort 6", each point labelled with its benefit-per-week ratio, and the top four by ratio accented. A footnote line: "Benefit and effort correlate at 0.91, so the quadrants largely restate a ranking. Order within each quadrant is set by benefit per person-week." Title: "Start the four highest-return projects: 302 of benefit for 23 person-weeks, at 13.1 per week against 9.9 for the big bets."

Read it. The 13.1304 against 9.8936 comparison is the decision, and it holds only if person-weeks are the binding resource. If the constraint were calendar time instead of staffing, carrier consolidation's 205 of benefit would dominate whatever its ratio, because ratio ranking assumes you can spend the freed weeks on the next project in the list. The correlation of 0.9095 is computed on nine points, so don't read it as an estimate of anything beyond this list.

Five things go on the checklist for a two-by-two. Print the cut rule on the axes. Check the correlation between the axes before claiming two dimensions. Publish the within-quadrant ranking. Name quadrants you would defend in a sentence about a specific project. And decide from the ranking, then draw the matrix to explain the decision.

Worked example 4: Repairing a concept diagram with Gestalt (India, elements)

Aarohan Credit is a synthetic lending marketplace. Its architecture slide has nine boxes: borrower, mobile app, credit bureau, underwriting engine, three lending partners and two collection agents. The drawing carries 9 boxes, 14 arrows, 9 icons, 3 stray labels, a legend, 9 drop shadows and a title, and the message it has to carry is that the marketplace originates and underwrites while all balance-sheet risk rests with the three partners. Within-group spacing is 14 points and between-group spacing is 16.

Solution. Step 1, count what is on the page. 9 + 14 + 9 + 3 + 1 + 9 + 1 = 46 objects.

Step 2, count the relationships the form can express, and the ones asserted. Nine nodes admit 9 × 8 / 2 = 36 undirected pairs. The drawing asserts 14 of them, which is 14 relationships the reader must inspect and hold.

Step 3, measure the grouping. The proximity ratio is between-group spacing over within-group spacing, 16.0/14.0 = 1.1429. Below about 2 the eye cannot resolve groups, so the reader is grouping from the labels, which is the work the diagram was drawn to save.

Step 4, rebuild. Keep the nine boxes and the title. Cut the arrows to the five that carry the claim: borrower to app, app to underwriting engine, bureau to underwriting engine, underwriting engine to partners, partners to collection agents. Replace the nine deleted arrows with three light enclosures labelled "Aarohan operates", "Partner balance sheets" and "Recovery". Delete every icon, stray label, the legend and all drop shadows. Tighten within-group spacing to 8 points and open between-group spacing to 32.

Step 5, measure the result. Objects: 9 boxes + 5 arrows + 1 title + 3 enclosures = 18, a reduction of 28 objects or 60.8696 percent. Proximity ratio 32.0/8.0 = 4.0000. Relationships asserted: 5 arrows plus 3 group memberships.

The wrong version. The 46-object original, whose specific failure is that it asserts every relationship the architect knows. It also fails the seven-rule check twice over. There's no single message, since three claims compete, and no colour restraint, since the icons and shadows carry no information. Its composite of 0.4580 misses the 0.80 pass mark by 0.3420.

The corrected version. Three enclosed groups reading left to right, five labelled arrows, one accent on the enclosure named "Partner balance sheets", and the title "All credit risk sits with the three lending partners; Aarohan originates and underwrites." A reader who reads only the title and the accented enclosure has the claim.

Read it. Nothing true was deleted. The nine removed arrows described real relationships that didn't carry this claim. They belong on an appendix architecture page for the engineering audience, a different reader with a different question. And the 60.8696 percent reduction is a property of this drawing. It isn't a target. A diagram with 12 objects can't be cut by 60 percent and shouldn't be.

The concept-diagram checklist starts with two counts. Count the objects, and count the arrows against the pairs the nodes admit. Measure the proximity ratio and get it above 2, ideally above 3. Replace arrows with enclosure wherever the relationship is membership rather than action. Delete every mark that encodes nothing, then read the title alone and check that it still lands.

Worked example 5: A six-stage cycle drawn three ways (US, days)

Halyard Systems, a synthetic enterprise software vendor, has measured its customers' buying cycle: awareness 21 days, research 34, shortlist 18, evaluation 45, approval 29, onboarding 26. The message is that the deal stalls in evaluation and approval, and the exhibit has to carry seven facts: six stage durations and the total.

Solution. Step 1, total and share the stages. The cycle runs 173 days. Awareness is 12.1387 percent, research 19.6532, shortlist 10.4046, evaluation 26.0116, approval 16.7630 and onboarding 15.0289. Evaluation is the longest single stage. Evaluation and approval together are 74 days, or 42.7746 percent of the cycle.

Step 2, count the objects in the version that arrived. A circular process wheel with 6 stage wedges, 6 stage labels, 18 sub-labels, 6 icons, 6 circular arrows, a centre ring, an outer ring and a legend: 45 objects for 7 facts, which is 6.4286 objects a fact.

Step 3, count the plain row. A row of six boxes with durations printed, joined by five arrows: 6 boxes, 6 labels, 6 duration labels, 5 arrows, 23 objects, or 3.2857 a fact. The wheel spends 22 more objects than the row, 95.6522 percent more, for the same seven facts.

Step 4, write it out. "The buying cycle runs 173 days from first awareness to onboarding. Evaluation and approval together take 74 of those days, 42.8 percent of the total, and they are where deals stall. Everything before evaluation takes 73 days combined and is largely outside our control." Three sentences, 44 words, zero objects.

The wrong version. The wheel, and its object count isn't even its worst property. It asserts that the process is a cycle, and a customer who onboards doesn't go back to awareness. It also makes duration comparison impossible. Six wedges of equal angle hide the fact that evaluation is two and a half times shortlist, and the sub-labels crowd the wedges until the type has to shrink.

The corrected version. Either the row, with the evaluation and approval boxes accented and the title "Evaluation and approval take 74 of the 173-day cycle: fix the middle, not the top of the funnel", or the three sentences on a text slide with the same title. If the deck already has a full page of prose, use the row. If it doesn't, the sentences win.

Read it. The three-sentence version carries an interpretation the drawings can't, which is that the first 73 days are largely outside the vendor's control. Neither drawing can say "largely outside our control", because that's a judgement rather than a duration. Part of the urge to reach for a diagram is an urge to avoid committing to judgements. The 42.7746 percent figure is elapsed time. It says nothing about where selling effort goes, which is a different measurement and would need its own exhibit.

Four questions go ahead of any sequence you're about to draw. Count the objects and the distinct facts, and compute objects per fact. Ask whether the shape asserts something the data doesn't support, such as a cycle that doesn't close. Write the three-sentence version first, and only draw if the drawing beats it. Where magnitudes matter, put the numbers in the labels instead of the geometry.

Worked example 6: Placing five process boxes by arithmetic (India, points)

Aarohan Credit's onboarding process has five steps, and the row has to sit on a 960 by 540 point widescreen canvas with a 48-point margin. The version that arrived was placed by dragging: left edges at 51, 228, 402, 585 and 761, widths of 150, 156, 152, 149 and 155, tops at 214, 210, 216, 211 and 215.

Solution. Step 1, define the grid once. Content width is 960 − 2 × 48 = 864. A twelve-column grid with a 24-point gutter gives a column width of (864 − 11 × 24)/12 = 50.0000 points and a pitch of 74.0000. Check: 12 × 50 + 11 × 24 = 864.0000.

Step 2, compute the five boxes. With a 24-point gap, width = (864 − 4 × 24)/5 = 153.6000 points. Left edges: 48.0, 225.6, 403.2, 580.8, 758.4. Right edges: 201.6, 379.2, 556.8, 734.4, 912.0. The last right edge meets the content boundary at 912.0000, a residual of 0.000000. A 90-point box centred vertically has its top at (540 − 90)/2 = 225.0000.

Step 3, measure what dragging produced. Gaps between boxes are 27, 18, 31 and 27 points: mean 25.7500, range 13, sample standard deviation 5.5000. Tops range over 6 points with a standard deviation of 2.5884. Widths range over 7 points with a standard deviation of 3.0496. The worst single left-edge error against the computed layout is 4.2000 points.

Step 4, count what the eye must absorb. The dragged version presents 3 distinct gap values, 5 distinct widths and 5 distinct tops: 13 values. The computed version presents 1 of each: 3 values.

The wrong version. Five boxes that look approximately right and are wrong in twelve small ways, none of which a viewer can name. The reader comes away with a vague sense that the slide is unfinished, and they'll put it down to the content. The gap sequence 27, 18, 31, 27 is the worst of it, because unequal gaps are read as grouping. An 18-point gap between boxes two and three says those two belong together, and the process makes no such claim.

The corrected version. Five boxes of 153.6000 points at lefts 48.0, 225.6, 403.2, 580.8 and 758.4, all with tops at 225.0000, all 90 points tall, with 24-point gaps throughout. Where five boxes won't hold their labels at full type size, the two-row variant on the same grid gives a top row of three at (864 − 2 × 24)/3 = 272.0000 points a box, nearly double the width.

Read it. No individual error in the dragged version exceeds 4.2 points, which is about a millimetre and a half on a projected slide. Alignment failures are invisible one at a time and cumulative in effect, so they get blamed on the content. The computed layout is cheaper to read, and the saving is ten judgements the reader doesn't have to make before reaching your content.

Placing diagram elements has its own short checklist. Compute the content width, then compute the box width from the count and the gap instead of choosing it. Check that the last right edge meets the content boundary with a zero residual. Use one gap value, one width and one top. If labels don't fit, change the layout. Leave the type size alone.

Worked example 7: One accent, measured by area and by contrast (US, ratio)

The Kestrel Supply process diagram has 13 shapes: five large boxes of 153.6 by 90 points, five medium of 120 by 72, and three small of 100 by 60. The version that arrived used seven hues, one per functional area. The message needs a single thing to stand out, the exception branch.

Solution. Step 1, compute the areas. Large 5 × 153.6 × 90 = 69,120.0 square points; medium 5 × 120 × 72 = 43,200.0; small 3 × 100 × 60 = 18,000.0. Total 130,320.0.

Step 2, size the accent two ways. Accenting one large box is 1 of 13 shapes, 7.6923 percent by count, and 13,824.0 of 130,320.0 square points, 10.6077 percent by area. Both fall in the range where an accent still reads as an accent, before it starts to look like a second palette.

Step 3, size the version being replaced. Seven hues across 13 shapes means no hue can claim more than 100/7 = 14.2857 percent of area, and none of them marks anything, because every shape is coloured. Colour has turned into a property of the diagram. It's no longer a signal within it.

Step 4, check contrast. Use the WCAG 2.2 normal-text threshold of 4.5 to 1 (checked September 2026; confirm, as guideline versions advance). Accent #B34700 on white gives 5.5002 to 1 and passes. White text on #B34700 gives the same 5.5002 and passes, so the accent works as fill and as ground. Neutral outline #6B7280 on white gives 4.8345 and passes. Body label #1F2937 on white gives 14.6791. And the light grey #9CA3AF that authors reach for to make things recede gives 2.5388 and fails.

Step 5, check the pair that actually matters. Accent against neutral, #B34700 on #6B7280, gives 1.1377 to 1. Those two differ almost entirely in hue and barely in luminance, so a colour-blind reader or a washed-out projector may not separate them at all.

The wrong version. Seven hues, one per functional area, with the exception branch in the same weight of colour as everything else. It looks designed. It marks nothing, and the light grey used for the supporting boxes is illegible to part of the room at 2.5388 to 1, so the diagram manages to be too colourful and too faint at the same time.

The corrected version. Everything in #1F2937 type on white with #6B7280 outlines, and one large box, the exception branch, filled #B34700 with white labels. A heavier outline weight reinforces the accent, and so does the branch annotation right beside it, so the emphasis survives a monochrome print.

Read it. The 1.1377 ratio is the finding here, and it applies well beyond this palette. Two colours can look obviously different on your monitor and still be nearly identical in luminance. So the rule is to encode twice. The contrast figures are computed from the WCAG relative-luminance formula and describe these particular hex values on a white ground. A slide on a dark template needs the whole set recomputed, and an accent that passes on white can fail on dark.

Colour in a diagram gets the shortest checklist of the seven. One accent, sized between roughly 5 and 15 percent of area, with everything else neutral. Compute the accent's contrast against the ground and against the neutral. Never let colour be the only encoding. Recompute the whole set if the template ground changes.

Verify

Every figure printed above, restated at the precision the computation produced. Reproduce them before you trust anything derived from them.

Practice set

Fourteen problems, alternating markets. Work each one before you read its solution. The arithmetic is small enough to do on paper, and the point of doing it is that the checks become reflexes. A calculator is assumed. Numeric answers within ±0.5 percent of the stated figure count as correct unless a problem says otherwise, except where the answer is a residual, which must be exact.

P1 (guided). Close a goods flow (India). A rice mill dispatched 1,240.0 tonnes in a month. The diagram shows 815.0 tonnes to the depot network and 402.0 tonnes direct to institutional customers. Someone has drawn only those two arrows. State the conservation identity, compute what is missing, and say what the diagram must show.

Solution. The identity is that what leaves the mill arrives, is consumed, or is lost: dispatched = depot + direct + losses. Depot plus direct is 815.0 + 402.0 = 1,217.0, so the residual is 1,240.0 − 1,217.0 = 23.0 tonnes. That 23.0 is shrinkage, and it has to appear on the diagram as a labelled terminal node. It can't just be absorbed, because a two-arrow diagram whose arrows don't sum to the source is claiming that nothing is lost. With the shrinkage node added, 815.0 + 402.0 + 23.0 = 1,240.0 and the residual is 0.0. Print the residual on the page. A reader who sees "residual 0.0 tonnes" knows the drawing has been checked. A reader who sees nothing has to trust you.

P2 (guided). Recover a missing net position (US). Five banks settle on a net basis. Four positions are published, in $m: A −18.4, B +7.2, C +25.6, D −9.3. The fifth, E, has been cut off the slide. Find it, then compute the settlement funding requirement, and state the property that let you do this.

Solution. Net positions across a closed system sum to zero, because every payment is one bank's debit and another's credit. The four known positions sum to −18.4 + 7.2 + 25.6 − 9.3 = +5.1000, so E must be −5.1000. Settlement funding is half the sum of the absolute positions, because the debtors' total equals the creditors' total and only one side needs funding. That gives (18.4 + 7.2 + 25.6 + 9.3 + 5.1)/2 = 65.6/2 = 32.8000 $m. The property is conservation, and notice that it also works in reverse as an error detector. If all five had been printed and summed to anything other than zero, at least one would be wrong.

P3 (guided). Weight a two-branch process (India). A KYC verification process resolves 62 percent of applications in 5 steps and sends the remaining 38 percent down a manual-review path of 12 steps. Compute the weighted average step count, and state why the weighted average is the wrong number to put on the slide.

Solution. 0.62 × 5 + 0.38 × 12 = 3.10 + 4.56 = 7.6600 steps. So why not print it? No application takes 7.66 steps. Sixty-two percent take five and thirty-eight percent take twelve, and the interesting fact is the 38 percent, more than a third of applicants going through a path more than twice as long. A weighted average on a process slide reports a state that nothing in the process occupies, and it hides the branch that carries the cost. Print the diamond with 62 percent and 38 percent on its exits, and put the average in a footnote if a reader wants it for capacity planning.

P4. Compute a four-box layout (US). A widescreen canvas is 960 by 540 points with a 48-point margin. Four equal boxes are to sit in a row with a 32-point gap. Give the box width and the four left edges, and state the check that closes the layout.

Solution. Content width is 960 − 96 = 864. Width = (864 − 3 × 32)/4 = (864 − 96)/4 = 192.0000 points. Left edges are 48.0, 272.0, 496.0 and 720.0, since the pitch is 192 + 32 = 224. The check is that the last box's right edge, 720.0 + 192.0 = 912.0, equals the margin plus the content width, 48 + 864 = 912, for a residual of zero. Any layout arithmetic that doesn't land on the content boundary has an error in it. Finding that error takes ten seconds, against the twenty minutes of nudging it would otherwise cause.

P5. Diagnose three spacings (India). Three versions of a grouped diagram have within-group and between-group spacings, in points, of (10, 12), (6, 24) and (9, 30). Compute the proximity ratio for each and say which will read as groups.

Solution. Ratios are 12/10 = 1.2000, 24/6 = 4.0000 and 30/9 = 3.3333. The first doesn't group. At 1.2 the eye can't tell the between-group gap from the within-group gap, so the reader falls back on reading the labels, which is the work the grouping was meant to save. The second and third both group cleanly, since anything at or above about 2 reads as separation and 3 or more reads as separation without effort. Notice that the absolute numbers don't decide this. Version three has the widest gaps of the three and still isn't the strongest grouping, because grouping is a ratio judgement rather than a distance judgement.

P6. Place seven initiatives in a two-by-two (US). Seven initiatives are scored out of 10 for impact and then cost: A (8,3), B (2,9), C (6,6), D (9,8), E (4,2), F (7,7), G (3,4). Using a median split with high impact meaning at or above the median and low cost meaning at or below it, compute both medians and assign each initiative to a quadrant.

Solution. Impact values sorted are 2, 3, 4, 6, 7, 8, 9, giving a median of 6. Cost values sorted are 2, 3, 4, 6, 7, 8, 9, giving a median of 6. A (8,3) is high impact and low cost: do now. B (2,9) is low and high: drop. C (6,6) is right on both medians, and under the stated at-or-above and at-or-below rule it goes in do now. D (9,8) is a big bet. E (4,2) is a fill-in. F (7,7) is a big bet. G (3,4) is a fill-in. C is the item worth arguing about. It's on both boundaries, its quadrant is entirely an artefact of the tie-breaking rule, and a matrix that doesn't print its rule has decided C's fate invisibly.

P7. Budget the elements (India). A concept diagram carries 41 objects. The instruction is to rebuild it at half the element count. Compute the target, state whether a rebuild to 17 objects meets it, and give the percentage cut.

Solution. Half of 41 is 20.5, so the target is 20 objects or fewer. A rebuild to 17 meets it with room to spare. The cut is (41 − 17)/41 = 58.5366 percent. The instruction is a forcing constraint, and it says nothing about the right number of objects. Its value is that it makes you rank the elements. Ranking is how you discover that nine of your fourteen arrows were relationships you found, ones the claim never needed.

P8. Check a teal palette (US). A diagram uses #0F766E as its accent, #D1D5DB for recessive boxes, and white labels on the accent fill. Compute the three contrast ratios against a white ground and state which fail the 4.5 to 1 threshold for normal text.

Solution. #0F766E on #FFFFFF gives 5.4733 to 1 and passes. #FFFFFF on #0F766E gives the same 5.4733 to 1, since the ratio is symmetric, so the accent works as a fill with white text on it. #D1D5DB on #FFFFFF gives 1.4735 to 1 and fails badly. That light grey is nearly invisible on white, so any label or outline drawn in it is unreadable for part of the audience, and at any projector brightness below ideal. The repair is a darker neutral around #6B7280, which measures 4.8345 to 1 on white. Recessive doesn't mean invisible.

P9. Price a wheel against a row (India). A five-stage circular diagram carries 52 objects and communicates 6 facts. Compute its objects-per-fact figure, compute the same for a five-box row with four arrows and five labels, and give the reduction.

Solution. 52/6 = 8.6667 objects a fact. The row is 5 + 4 + 5 = 14 objects, so 14/6 = 2.3333 a fact, and the reduction is (52 − 14)/52 = 73.0769 percent. Above about four objects a fact, ask what the extras buy. Here they buy a ring, a legend and a set of icons, none of which carries a fact, plus the false claim that a five-stage sequence returns to its start.

P10. Score three slides (US). The composite is 0.6 × machine + 0.4 × rubric, and the pass mark for a non-crown node is 0.80. Score these three and say which pass: machine 0.71 with rubric 0.75; machine 0.88 with rubric 0.65; machine 0.95 with rubric 0.95.

Solution. 0.6(0.71) + 0.4(0.75) = 0.426 + 0.300 = 0.7260, which fails. 0.6(0.88) + 0.4(0.65) = 0.528 + 0.260 = 0.7880, which fails. 0.6(0.95) + 0.4(0.95) = 0.9500, which passes. The second is the instructive one. Near-clean mechanics with a weak self-assessed argument still misses, by 0.012. That's the weighting doing its job, because a diagram that's well drawn and says nothing is the failure all of this is meant to prevent.

P11. Count swimlane risk (India). For 3, 4 and 6 lanes in a strict relay, give the number of handoffs and the maximum number of lane-to-lane transitions the notation permits if any lane can call any other. Then say what the second number is for.

Solution. A strict relay through n lanes has n − 1 handoffs: 2, 3 and 5. If any lane can call any other, the ceiling is n(n − 1) ordered pairs: 6, 12 and 30. The second number is a design warning. It doesn't count anything real. It says that as lanes grow, the notation's capacity for crossings grows quadratically while the reader's capacity to follow them doesn't. So a six-lane diagram usually reads as a tangle, and the better fix is often two diagrams, one per phase.

P12. Compute netting efficiency (US). A clearing system moves $412.5m of gross payments in a day and requires $37.8m of settlement funding. Compute netting efficiency and state what the figure does and does not claim.

Solution. 1 − 37.8/412.5 = 1 − 0.091636 = 0.908364, or 90.8364 percent. It claims that 90.8 percent of the day's gross value was extinguished by offsetting positions, with no cash moving. It doesn't claim that the system is safe. A high netting efficiency means participants owe less at settlement, and it also means each participant's obligation depends on every other participant performing. That concentrates risk instead of reducing it. Efficiency and resilience are different measurements, and a diagram that prints one shouldn't be captioned as though it printed the other.

P13. Test the cut rule (India). Seven initiatives have annual benefits, in ₹ lakh, of 120, 35, 60, 240, 45, 52 and 28. Compute the median and the mean, and identify which initiatives are classified as high benefit under one rule and low under the other.

Solution. Sorted: 28, 35, 45, 52, 60, 120, 240. The median is 52 and the mean is 82.8571, pulled upward by the single 240. Under a median split, 52, 60, 120 and 240 are high benefit. Under a mean split, only 120 and 240 are. The two initiatives at 60 and 52 flip from high to low on nothing but the choice of statistic. That's two of seven. When one item is several times the others, the mean stops describing the middle, and a median split is usually the defensible choice. Either way, the decision must be printed.

P14. Find the missing arrow (US). A claims-handling flow shows $5,400,000 of claims received in a quarter, and three exits: $3,950,000 paid, $820,000 held pending documents and $615,000 written off. Compute the residual and say what to do about it.

Solution. The exits sum to 3,950,000 + 820,000 + 615,000 = 5,385,000, against 5,400,000 received, so the residual is $15,000. Don't adjust a number to close it. A residual means one of three things. An exit exists that isn't drawn, such as claims transferred to another handler. Or a figure was transcribed wrongly. Or the period boundaries differ between the source tables, which is the commonest cause and the easiest to miss. Find which one it is, then either draw the missing exit or fix the number, and print the residual as zero afterwards. If a diagram closes to $15,000 of nothing, a reader is entitled to doubt its other numbers.

Applied mini-project

Take one mechanism you already understand from anywhere in the course and turn it into a diagram. Budget about 150 minutes. Suitable mechanisms include a cash conversion cycle, a securities settlement chain, a dividend imputation flow, an options assignment process, a fund's subscription and redemption path, a credit approval workflow, or a supply chain you have studied. Pick something you can already explain out loud. The exercise is about the drawing, so the mechanism should already be familiar.

Step 1: write the message and the reader, before drawing (about 15 minutes). One sentence that a reader could disagree with, plus one named role who would act on it. Then state which diagram family the message implies and why the other four don't fit. If your message is a magnitude comparison, stop. Choose a different mechanism. The right answer there is a chart, and this project won't teach you anything.

Step 2: draw the complete version (about 40 minutes). Every node, every relationship, every branch, drawn correctly under the notation in sections 3 and 4. This version is for you. Its job is to prove you understand the mechanism. For once the failure to worry about is leaving things out, so include the awkward parts.

Step 3: run the checks (about 20 minutes). State the conservation identity if the diagram carries a quantity, and print the residual. Count the objects, count the distinct facts and compute objects per fact. Count the nodes, compute the pairs they admit, and count the relationships you asserted. Measure the proximity ratio. Compute the box widths and gaps from the canvas and record them.

Step 4: draw the argument version at half the elements (about 45 minutes). Rebuild, keeping only what carries the message from step 1, with a target of half the object count of step 2. Name the Gestalt principle that carries the new version: proximity, similarity, enclosure, continuity or closure, and say in one sentence how it replaced the elements you deleted. Add one accent, sized between roughly 5 and 15 percent of shape area, and compute its contrast against the ground and against the neutral. Write the title as a full sentence carrying the finding.

Step 5: write the deletion log and the stated limits (about 30 minutes). List every element you removed and, for each, the relationship it expressed and why the message doesn't need it. Then write three sentences on what the diagram can't show, including at least one magnitude the reader might wrongly infer from the geometry.

Artefacts and their tests.

1. The message sheet: one sentence, one named reader, the family chosen and four families rejected with reasons. Test 1: the sentence is disputable, and a reader could hold the opposite position without being obviously wrong.
2. The complete diagram. Test 2: the notation is correct, meaning every branch is a diamond with labelled exits carrying shares, every loop returns to the specific step repeated, and every arrow carries a verb.
3. The check sheet: conservation identity and residual, object count, fact count, objects per fact, node count, pairs admitted, relationships asserted, proximity ratio, computed box widths and gaps. Test 3: the residual is exactly zero or equals a named, printed reconciling item, and the layout arithmetic meets the content boundary with a zero residual.
4. The argument diagram. Test 4: object count at or below half of the complete version, proximity ratio at or above 2.0, exactly one accent between 5 and 15 percent of shape area, accent contrast at or above 4.5 to 1 against the ground, and a second non-colour encoding on the accented element.
5. The title. Test 5: a full sentence with a verb, carrying the finding and, where the data supports one, a number.
6. The deletion log and limits. Test 6: every deleted element accounted for, and at least one stated limit that a reader could otherwise have inferred wrongly from the drawing.

Rubric, 24 points, with a pass at 19 and no zero permitted on any row.

Any artefact whose test fails scores zero on its row whatever the drawing looks like, and a diagram whose residual doesn't close scores zero on row 3 even if every other row is a three. The reason is the one you've met all the way through. A diagram is a claim, and a claim that fails its own arithmetic isn't improved by being beautiful.

Reading & resources

Core (do these).

• Edward Tufte, The Visual Display of Quantitative Information: the origin of the data-ink argument, which is the ancestor of the element counting used here. Read it for the standard of care. Several of its specific prescriptions have been argued with since. [Paid] [Intermediate]
• Stephen Few, Show Me the Numbers: the reference treatment of table and graph design, and the clearest published account of why grouping and alignment change reading speed as well as appearance. [Paid] [Intermediate]
• Colin Ware, Visual Thinking for Design: the perception science under the Gestalt rules, written for people who make things. The chapters on pattern and structure are the ones that pay back immediately. [Paid] [Advanced]
• Barbara Minto, The Pyramid Principle: the hierarchy diagram done properly, and the source of the grouping and ordering rules that keep a tree from overlapping at its second level. [Paid] [Intermediate]
• Gene Zelazny, Say It With Charts: the message-to-form mapping, written for a consulting audience. Its treatment of concept and process visuals is the part that matters here, since the chart material belongs to the chart-craft branch. [Paid] [Beginner]

Going deeper.

• Robert Horn, Visual Language: an ambitious and unusual treatment of diagrams as a language with a grammar, which is the framing used here. It's uneven, and worth reading for the framing more than for every claim. [Paid] [Advanced]
• Alberto Cairo, How Charts Lie: mostly about charts, and the chapter on how visual form makes claims transfers directly to diagram geometry that asserts a cycle, a funnel or a hierarchy the data does not support. [Paid] [Beginner]
• Dan Roam, The Back of the Napkin: the case for drawing to think rather than drawing to present. Read the first half. The taxonomy in the second half is less useful than the habit the first half installs. [Paid] [Beginner]
• Jean-luc Doumont, Trees, Maps, and Theorems: compressed and opinionated, and the best short account of why a graphical display should carry a single message. [Paid] [Advanced]
• Bang Wong's series on points of view in Nature Methods, 2010 to 2011: short, free-to-find columns on colour, salience and encoding, written for scientific figures and applying without modification to business diagrams. [Free] [Intermediate]

Standards and tools (free).

• W3C Web Content Accessibility Guidelines: the source of the 4.5 to 1 contrast minimum for normal-size text and 3 to 1 for large text (WCAG 2.2, checked September 2026; confirm, since guideline versions advance and the thresholds can move).
• Business Process Model and Notation (BPMN), from the Object Management Group: the formal notation behind the step, decision, lane and event marks used informally here. You don't need the full specification, and reading its treatment of gateways once will make your decision nodes noticeably more precise. [Free] [Advanced]

Do this, not just read. Take the last process or architecture diagram you received from someone else. Count its objects and the distinct facts it carries, then divide, which gives you its objects-per-fact figure. Now count its nodes and work out how many pairs they admit, against how many relationships the drawing actually asserts. Two numbers, four minutes, and you'll know more about that diagram than its author does.

Teach it back & journal

Feynman prompt. Explain to a colleague who's never thought about diagrams why two pictures carrying the same information can differ by a factor of twenty in what they cost a reader. Do it without using the words clean, professional or visual. Use the arithmetic. Twelve payment legs among four banks give a reader twenty mental operations to find the funder, or one lookup if the net is printed on the node. Nine nodes admit 36 relationships, and a drawing that asserts 14 of them has handed the reader 14 inspections, of which the claim needed 5. A six-stage wheel spends 45 objects on 7 facts where a row spends 23 and three sentences spend none. Then make the harder point, which is that none of those numbers is about drawing. Each one is about a decision taken before anything was drawn: what the reader must understand, and therefore what can be left out. If your colleague can tell you, unprompted, why a flow diagram needs a printed residual, you've taught the idea, not just the arithmetic.

Journal prompt. Find a diagram somebody sent you in the last month, ideally an architecture picture or a process map you were expected to understand without help. Run four measurements on it and write the numbers down. Count the objects and the distinct facts, and divide. Count the nodes, compute the pairs they admit, and count the relationships the drawing asserts. Estimate the within-group and between-group spacing and take the ratio. Find the conservation identity if one exists, and see whether the quantities close. Then write a paragraph on what the author was optimising for, using your own evidence instead of a guess. A diagram with 14 arrows among 9 nodes was optimising for completeness, one with seven hues for looking finished, and one with a wheel over a straight sequence for the appearance of a system. None of those authors was careless, and that's the uncomfortable part. Each of them would have drawn better with a message written first. Finish by taking one diagram of your own, rebuilding it at half the element count, and recording what you were most reluctant to delete and whether the rebuilt version actually suffered for losing it. That last note is the one worth keeping. The elements you protect are a stable feature of how you think, and knowing yours is worth more than any rule here.