The Ghost Deck and Dot-Dash: Sketch the Story Before You Build

Presentation track, branch pr-visual (Diagrams & Storyboards) · Node PR3.02 · ~8 focused hours · Mastery gate ≥ 85%

An analyst builds fourteen pages over nine hours. A colleague reads them for half an hour and cuts five. Two of the nine survivors come back with the wrong chart on them and have to be rebuilt. Total elapsed time, eleven and a half hours, of which two hundred and fifty minutes produced nothing that reached a reader.

Nobody in that story did bad work. The five cut pages were accurate, aligned and correctly sourced. They were cut because the argument did not need them, and the argument did not need them for reasons that were visible on the day the deck was scoped. What made the waste expensive was not the cutting. It was that the cutting happened after the building.

Run the same deck the other way. Fourteen sheets of paper, four minutes each, one headline and a rough box for the exhibit: fifty-six minutes. The same half-hour review cuts the same five pages, and because the sketches show the chart forms, it also catches one of the two form errors before anything is built. Build the nine survivors, rebuild the one exhibit the review missed, and the deck ships in eight hours and sixteen minutes. The saving is three hours and fourteen minutes on a single deck, which is twenty-eight percent of the first path.

That arithmetic is the whole case, and the interesting part of it is how few pages have to be wrong for it to work. At the build rates above, the ghost pass pays for itself the moment the review cuts more than 1.44 pages. Cut one page and you are seventeen minutes down. Cut two and you are twenty-two minutes up. Nobody has ever scoped a deck that survived its first review without losing two pages.

Paper is not the point either. The point is that a page which costs four minutes gets killed on its merits, and a page which cost fifty-five minutes gets defended. That is a fact about people, not about stationery, and it does not go away when the sketching happens in a text file or on a tablet. The sunk cost has to be small enough that you can throw the page away without arguing for it, and small enough that you will draw the alternative rather than talk yourself into the version you already have.

What follows is a working method with two named parts. The ghost deck is one page per slide, drawn before any real page exists, carrying a headline and a sketch of the evidence. Dot-dash is the discipline inside each page: the dot is the claim the page makes, in fourteen words or fewer, with the number that carries it; the dash is the single line of evidence that proves the dot, naming the exhibit, the figure and where the figure comes from. A ghost deck without dot-dash is a list of topics with boxes drawn round them, which is the failure mode most people arrive with.

The refusal at the end is the part learners find hardest and the part that pays best. You do not open the software until the dots hold together on their own. Not because tools are bad, but because the tool answers questions about appearance and the ghost deck answers questions about argument, and appearance work performed on a wrong argument is the most convincing waste in professional life.

Learning objectives

You can:

1. Compute the rework cost of building before sketching, from stated page counts, per-page build times and a cut rate, and state the break-even number of cut pages at which a ghost pass pays for itself.
2. Build a ghost deck from a finished analysis: one page per slide, each carrying a dot, a dash, an exhibit sketch, a source and a status, with the page count fixed before the first page is drawn.
3. Write a dot that passes four tests: it is a claim, it carries the number that decides it, it runs to fourteen words or fewer, and it is not a restatement of the dot above it.
4. Write a dash that names the exhibit form, the specific figure and the source, and recognise the dash that is secretly two dashes, which means the page is secretly two pages.
5. Sketch an exhibit in about ninety seconds in three candidate forms, recording the axes, the sort order and the series that carries the accent, and choose between the three against the dot rather than against taste.
6. Run the read-only-the-dots test on a ghost deck, score it against a fixed question set, and diagnose the three failure signatures: the label dot, the duplicate dot and the leap between dots.
7. Cut pages on paper using a stated ceiling, defend each cut with the question it fails to answer, and compute what the cut saved in build minutes.
8. Set a page count from airtime, dividing the speaking minutes left after a discussion reserve by the minutes per page the format needs, and state what the resulting number implies about the audience.
9. Apply the placeholder discipline where the data does not exist yet: bound the unknown with an honest range computed from what is disclosed, mark its status, and never let a placeholder cross into a shown page as a fact.
10. Convert a ghost deck into a build plan, ordered by data readiness and argument load rather than by page number, and state the condition under which the correct output of a ghost deck is a decision not to build at all.

Prerequisites & connections

Builds on. PR3.01 is the branch prerequisite. It supplies the visual vocabulary the exhibit sketches use: the five diagram families, the Gestalt grouping principles and the habit of drawing the mechanism instead of the org chart. PR0.01 supplies the three inputs a ghost deck consumes and doesn't re-derive, which are the one-sentence takeaway, the audience profile and the medium choice. Arrive without those three and the first page you draw will be a title page for a deck whose purpose nobody has stated. M8.02, or any completed case, gives you an analysis of your own to storyboard. That makes the practice worth roughly double what a supplied brief is worth.

Feeds forward. PR4.01 takes the ordered dots and gives them a formal structure: the governing thought, the key line, and the situation-complication-question-answer opening. PR4.02 owns the title stack as a sequence and the horizontal and vertical logic tests that a finished deck must pass. PR2.01 through PR2.03 own everything about the exhibit that the sketch leaves undecided on purpose, which is form choice in full, decluttering, colour and the action title. PR5.01 to PR5.03 take a validated ghost deck and pour it into an archetype template. PR7.01 runs the sequence end to end under a time limit, and most of the time saved in the capstone comes from the ghost pass.

Five neighbours own material that touches this one, and none of it is rebuilt here. CN1.02 owns the ghost deck as a consulting planning tool: reverse-engineering a lean workplan from a day-one hypothesis, with the guardrails that keep placeholder numbers from anchoring a team. Here the same artefact does a different job. You use it to compress and test a story you've already finished, where CN1.02 uses it to scope analysis you haven't run yet. CN1.03 owns answer-first synthesis, the governing thought and the so-what ladder, and its own hand-off note draws the boundary from the consulting side: claims and their order belong there, slides and their look belong here. M10.03 owns the written analyst deliverables, the memo, the report and the flip test applied to a finished title stack. PR4.01 owns the pyramid and the SCQA opening as structures, while the sketching practice that comes before building belongs here. PR2.01 owns chart-form choice. The compact grammar table below exists only so a sketch can name a form correctly, and it's shallow on purpose.

One boundary is worth stating twice because learners cross it by accident. A ghost deck isn't a draft deck. A draft deck is a real deck that isn't finished. A ghost deck is a different artefact with a different unit of work, and the moment you start adjusting the spacing on a sketch you've stopped doing the cheap thing and started doing the expensive thing badly.

Core teaching content

1. Why the tool is the wrong place to think

Do the arithmetic properly first. The case for sketching usually gets made as an aesthetic preference, and it's really a cost calculation.

Take a deck of fourteen first-pass pages: eight exhibit pages, four text pages and two section dividers. Assume an exhibit page takes 55 minutes to build in the tool from data that already exists, a text page takes 25 minutes, and a divider takes 5. Argue with these numbers before you accept them. Building all fourteen costs 550 minutes. A colleague then reads the deck for 30 minutes and cuts five pages, three exhibits, one text page and one divider. They also send two exhibits back because the chart form doesn't carry the claim. Rebuilding those two costs 110 minutes.

Path B costs 194 minutes less, which is 28.1% of Path A. Look at where the saving comes from, because building faster has nothing to do with it. The five cut pages carried 195 minutes of build work that produced nothing. The extra rebuild in Path A carried another 55. Together that's 250 minutes of Path A, 36.2% of the total, spent on pages a reader never saw. The ghost pass costs 56 minutes and buys back 250, and the difference, 194 minutes, is the saving. Hold on to that identity. Sketching doesn't make you faster. It makes fewer of your minutes land on pages that die.

If a sceptical colleague pushes back, the break-even is the number that settles it. The average cut page carried 39 minutes of build work, so the 56-minute ghost pass pays for itself at 1.44 cut pages.

Look at the second row. By the time the review cuts a second page, the ghost pass is already ahead.

Price it at a blended analyst rate, and replace the rate with your own. At $95 an hour Path A costs $1,092.50 and Path B costs $785.33, a saving of $307.17 on one deck. At ₹2,400 an hour the same deck costs ₹27,600 and ₹19,840, a saving of ₹7,760. Both rates are stated assumptions, and neither comes from a survey. It's the ratio that travels. Whatever the rate, the saving is 28.1% of the build-first path.

There's a second cost the table doesn't price, and in practice it's larger. A page you've built is a page you'll defend. In Path A the five pages were cut by a colleague reading a finished deck, in a conversation where the author has already spent nine hours and has opinions. In Path B the same five pages are cut from sheets that cost four minutes each, by an author who hasn't yet decided that page 7 is good. Nobody has measured the size of that effect in a way worth quoting, so treat it as a mechanism without a number attached. Sunk cost is a property of the person, and the ghost deck's job is to keep it small enough that judgement survives. You've probably defended a page like that yourself.

2. The ghost deck: one page per slide, before any pixels

A ghost deck is a physical or textual object with one entry per intended slide, drawn in sequence, before a single real page exists. Each entry carries five fields and no more.

Five fields, and a page with an empty field isn't a page yet. The commonest empty field is source. When the source line is blank, you believe the dot. You don't know it.

Don't spend the hour arguing about format. Index cards work, a sheet of A4 folded into eight works, a plain text file with five lines per page works, and a whiteboard with sticky notes works when three people are building the deck. A page has to be cheap, movable and disposable. Sticky notes buy the movability directly, so teams reach for them. A text file buys version history, and that matters more when someone in another city will review the ghost deck.

Two rules govern the whole artefact. First, one dot per page, so a page that carries two claims is two pages or one claim you haven't found yet. Second, you write the dots in the order the reader will meet them. That makes the ghost deck an ordered sequence, and reordering it is the cheapest structural edit you have.

Number the pages in pencil, or leave the numbers off entirely until the sequence stops moving. A numbered ghost deck starts to feel finished, and a ghost deck that feels finished has lost the property you built it for. Pencil numbers are an invitation to keep moving cards around.

3. Dot-dash: the claim and the line that proves it

The dot is the page's claim. The dash is the evidence. Writing them as a pair, on the same card, in that order, forces a question that a deck outline never asks: what specifically would make this page true?

Four tests tell a dot apart from a heading.

A dot is a claim, meaning a competent colleague could hold the opposite view. "Revenue and lab-count trends" is a subject. "Revenue grew 15.4% a year while revenue per lab fell 14%" is a claim, and someone could argue that the fall is a mix effect that doesn't matter.

A dot carries its number. The figure that decides the claim goes in the sentence itself, and not just underneath it in the exhibit. A dot without a number is usually a dot whose number hasn't been computed, so writing the dot doubles as a check on whether the analysis is actually finished.

A dot runs to fourteen words or fewer. The exact value of the cap is arbitrary. Its effect isn't. A sentence that fits in fourteen words has one claim in it. A sentence that runs to thirty has two claims and a qualifier, and the reader has to do the separating. Here's a real failure and its repair, both counted:

Failing dot, 16 words. "Labs opened in the last two years bill a third of what a mature lab bills." Tightened dot, 12 words. "Labs opened since FY24 bill a third of a mature lab's revenue."

The repair kept the number, kept the claim, and swapped a relative time expression for a date. That also helps a reader who meets the page six months later. On a first draft, compression of this kind runs about 2.4 times. A 26-word label headline, written by an analyst who was describing the page and not arguing it, becomes an 11-word dot.

And a dot doesn't restate the dot above it. Two adjacent pages making the same claim in different words is the second commonest defect in a first-pass ghost deck. Inside the tool you can't see it, because the two pages look different.

The dash is one line, and it does three jobs. It names the exhibit form, such as "sorted bar", "slope chart over FY22 to FY26" or "two-column table". It states the figure the exhibit will show. And it names the source. If a dash needs two lines to do those three jobs, it's telling you the page carries two claims. Split it and re-read the sequence, because one of the two halves is often a page you don't need.

Worked through on one page from the India example below, the pair reads:

Dot. Sahyadri's expansion is earning 4.6% on new capital against a 12.5% hurdle. Dash. Bar pair, incremental ROIC against WACC: ΔNOPAT ₹22.67 cr over ΔIC ₹493 cr = 4.60%, from the FY22 and FY26 balance sheets and the tax note. Status: computed.

Check the dash against its three jobs and you'll find all of them done in one line.

Everything a builder needs is on that card, and a reviewer can attack the claim without waiting for the chart. Notice what's missing. No colour, no font, no title styling, no decision about whether the bars are horizontal. Those belong to the build, and they're cheap once the claim is settled.

4. Sketching an exhibit: enough to know the form, not enough to fall in love

An exhibit sketch takes about ninety seconds and answers four questions: what is on each axis, what the sort order is, which single series carries the accent, and what number the dot is claiming. It doesn't answer any question about appearance.

The economics here are even more lopsided than the page-level case. Sketching one candidate form costs 1.5 minutes, so three cost 4.5 minutes, against 55 minutes to build one exhibit in the tool. For the price of a single built chart you can sketch about 36 candidate forms. Almost nobody needs 36. Everybody needs three. Why three and not one? The first form that occurs to you is usually the shape your data is stored in, when what you want is the shape your claim needs.

Sketch three, then choose against the dot. The choice follows a grammar that gets its full treatment in the chart-craft work. A sketch needs only the summary.

Notice how often the right-hand column says pie.

Three standing rules sit under the table. Apply them at sketch stage, because breaking them at build stage costs a rebuild. Sort by value unless the order is intrinsic. A fiscal year has an order, and a list of regions doesn't. When you're comparing two series with different units, prefer small multiples to two axes, because the reader can't see that a scaling choice was made. And keep one message per exhibit, which is the one-dot-per-page rule seen from the chart's side.

Ugliness is what makes a sketch useful. Draw it badly on purpose. A sketch that looks good is one you'll keep when the claim changes, and the claim changes about a third of the time between the ghost pass and the build. Pencil, thirty seconds, no ruler.

One test closes the sketch. Write the number the dot claims directly on the drawing, and check that the drawn shape would show it. Say the dot claims that new labs bill a third of mature labs, and you've sketched a line chart of total revenue over five years. It fails the test instantly, because that shape can't display the ratio claimed. Catch that on a card and it costs nothing. Catch it after the build and it costs 55 minutes and a conversation.

5. Pressure-testing the flow: reading only the dots

Lay the cards out in order, cover the dashes and the sketches, and read only the dots. Then hand the same list to someone who hasn't seen the analysis and ask them the questions the audience will actually ask. That's the read-only-the-dots test, and it's the highest-yield twenty minutes in the method.

Use a fixed question set, agreed before the reading, of about six questions drawn from the brief. For a capital-allocation read-out the six are: what is the recommendation, what is the evidence for it, what would change it, what does it cost, who owns the first step, and when does the decision get re-tested. Score one point per question the reader can answer from the dots alone. Pass at five of six, which is 83.3%. That's a high bar. Leave two of the six unanswered and the deck fails.

Two scored examples, both worked in full later:

Neither first pass was a bad deck. In both, the argument lived in the exhibits and in the presenter's head, and the headlines didn't carry it. That's the normal condition of a first draft, and it's what the test is built to expose.

Three failure signatures account for almost every miss.

The label dot is a page whose headline names a subject instead of making a claim. It's easy to spot in a list and nearly invisible on a slide, because a slide with a good chart under a label headline looks competent. In the Sahyadri first pass, four of the eleven dots were labels, and those four pages were carrying two of the six answers.

The duplicate dot is two pages making the same claim. It usually happens where an analysis produced two exhibits for one finding, and the author, unwilling to discard either, gave each a page and split the claim across them. The repair is to keep the exhibit that shows the number in the dot and demote the other to the appendix.

The leap is a gap between two adjacent dots that the reader has to jump. Page 4 says mature labs are healthy, page 5 says pause new lab additions, and nothing in between says why the health of the mature estate implies anything about new capital. A leap is the most expensive of the three, because rewriting a headline won't fix it. You fix it by inserting a page, which means the analysis has a hole in it. That hole is the finding you wanted the ghost pass to surface.

Run the test with a person if you can, and with a twelve-hour gap if you can't. Reading your own dots cold the next morning finds most of the leaps, because overnight you stop being able to supply the missing sentence from memory.

6. Cutting on paper, which costs minutes rather than hours

Cutting is the operation the whole method exists to make cheap, and it needs a rule. Ask "is this page worth keeping?" and every page passes when its own author is the one asking.

The rule that works is a ceiling, set before anyone reads the cards. Decide the page count first, using the arithmetic in the next section, then cut to it. A ceiling turns an unanswerable question, is this page good, into an answerable one, is this page in the top nine. People rank reliably. They judge absolute merit badly, and you'll see the difference in every first-pass ghost deck that has fourteen good pages and room for nine.

Rank against the six-question set from the read-only-the-dots test. Each page either answers one of the six questions, supports a page that does, or does neither. Pages in the third group go, in order, until the count is met. In the Sahyadri example the first pass ran to eleven pages against a nine-page ceiling, and the two cuts were a market-context page and a methodology page. The market-context page answered none of the six. The methodology page answered "what would change it" less well than an appendix note answering the same question, so the claim moved and the page went.

Record the reason with the cut. One line: which of the six questions the page failed to answer, or which surviving page already carried its claim. Two things depend on that line. Someone will ask on Thursday why the market-context page is gone. If you can say it answered none of the six questions the sponsor asked, you're in a different position from someone who says it felt weak. And now and then a cut page has to come back. The reason line tells you what would have to change for that to be right.

The saving on those two Sahyadri pages, priced at the same build rates, was 80 minutes of build work against the 8 minutes the two cards cost. Ten to one, on two pages, on one deck. Across a year of decks, that ratio adds up.

Keep the cut cards. Don't throw them away, and don't delete the lines. A cut list is the cheapest appendix index there is. When a director asks about market context on Thursday, you can say it was considered, it's on a card, and it was cut for a stated reason. That answer is worth more than the page would have been.

8. Placeholders, and the number nobody has disclosed

Some pages need a number that doesn't exist yet. The data is coming on Tuesday, the company doesn't break out the segment, or the finding depends on interviews that are still being scheduled. A ghost deck handles this well and a draft deck handles it badly. A draft deck has a chart on the page, and a chart has to contain something.

The rule has two halves. A placeholder is marked, and a placeholder is bounded. Marked means the status field says estimated or missing instead of computed, and the mark travels with the card. Bounded means you compute what the unknown could be from what's disclosed. You don't write a plausible figure and move on.

People skip the bounding, and it's the part that does the work. Sahyadri doesn't disclose EBITDA by lab cohort. It does disclose total FY26 EBITDA of ₹183.481 crore, and you can compute the cohort revenue split at ₹645 crore mature and ₹86 crore new. Assume a mature-lab margin, subtract, and the residual is the new-lab margin that assumption implies.

A few points of assumed mature margin swing the new-lab answer from a comfortable profit to a loss.

You might expect the placeholder to be a number. Here it's a sentence, saying that the new-lab cohort's EBITDA margin lies somewhere between −11.65% and +14.60%, a band 26.25 points wide, and that the width of the band is itself the finding. A 26-point band means the disclosure can't support a claim about new-lab profitability at all. A page that was going to make one has to be rewritten to make a different claim or cut.

Notice what else the bound does. It tells you what one question to the finance team is worth. Ask for the mature-cohort margin and the band collapses to a point. That's the highest-value data request in the whole engagement, and the ghost deck found it in four minutes.

There are three things a placeholder must never do. First, it must never appear on a page shown outside the team as though it were a fact. The consulting material on hypothesis-driven work states the same guardrail from its own side, and it applies here unchanged. Second, it must never be a round number chosen because it looks reasonable, because a reasonable-looking figure anchors the team that wrote it harder than an obviously fake one. Third, it must never survive into the build without its status field. The build is where a placeholder acquires axis labels, a source line and the appearance of having been measured.

Where nothing can be bounded, write the sentence the page would make and leave the exhibit box empty with the word missing in it. An empty box in a ghost deck is a legitimate object. An empty chart in a built deck isn't, and that difference makes the sketching stage the cheap place to be uncertain.

9. Reviewing a ghost deck with someone else

Review a ghost deck alone and you'll find most of the label dots and almost none of the leaps, because your own knowledge fills a leap automatically. Get a second reader. Ask them two specific questions. Anything vaguer gets you the reassurance that wastes a review.

The first question is "which of these pages could you not defend if someone pushed back?" Asking which page is weak invites politeness. Asking which page this reader couldn't personally defend in a room turns the review from a critique of your work into a test of their comfort, and people will tell you straight whether they're comfortable.

The second question is "after page 4, what do you expect page 5 to say?" Ask it before they read page 5. If their expectation matches your page, they've confirmed your sequence. If they expect something else, they've found either a leap or a better order, and both are worth more than a compliment. Don't run it at every joint. Pick two or three, the ones where you were least sure.

Tell the reviewer two things before they start. First, the page ceiling, so their suggestions land inside the real constraint and nobody adds a fifteenth page to a nine-page deck. Second, tell them in one line who the audience is and what they're deciding. A reviewer who doesn't know that defaults to reviewing the writing.

Don't walk them through it. A ghost deck read aloud by its author scores far better than the same deck read silently by its audience, for the obvious reason that you supply the connective tissue as you speak. Hand it over, say nothing, and take notes. If you can't resist explaining, leave the room.

Where the review is remote, a plain text ghost deck beats photographs of cards, and a numbered list of dots pasted into a message beats a file attachment, because the reviewer will read the message and may not open the file.

10. From ghost deck to build plan

A validated ghost deck converts into a build plan, and the plan isn't the page order. So why not build in page order? It's the default, and it's wrong for two reasons that compound.

Order the build by data readiness first. Pages whose status is computed are buildable now. Pages whose status is estimated or missing are blocked, and starting them produces a placeholder chart that will be rebuilt. Send the data requests before you build anything. Then build the ready pages while the requests are out, and the blocked pages arrive with their data instead of waiting for it.

Order by argument load second. Among the buildable pages, build the two or three that carry the recommendation before the supporting pages, because those are the pages most likely to change shape once they're real. An exhibit that looked right as a sketch and reads wrong at full size is a discovery. You want to make it on page 3 of the build, while there's still time for the sequence to absorb it, and not on page 12.

Build the furniture last. Title pages, section dividers, agenda pages, the appendix index and the source footers cost little and change with the sequence. A divider for a section that later merges into another is a small waste, but it's waste with no upside. Leaving furniture till last also stops it becoming the procrastination task it usually is.

The nine-page Sahyadri deck needs only five lines of build plan.

Read down the table and you'll see the pages come out of order on purpose.

Write one line at the top of the plan explicitly: the condition under which you stop and go back. State it as a trigger. If the cohort margin arrives outside the bounded band, pages 5 and 9 change and the sequence gets re-read before any further building happens. Without a stop condition, a new fact is just an inconvenience. With one, it's a finding.

11. When the ghost deck says the analysis is not finished

Sometimes a ghost pass shouldn't produce a deck at all. Its right output is the sentence "we cannot answer the question yet", delivered on the day the sketching happened and not the day before the meeting.

Four signals say so, and you can see all of them on the cards.

The clearest is a recommendation page with no dash. The last card says what should be done, and the evidence line under it is empty or circular. Nothing in the analysis supports the action. It came from judgement, which may be right and isn't yet an argument.

Second, a leap that no page you could build would fill. You know page 5 needs to explain why mature-estate health implies a pause on new capital. You draw the card, and you can't write a dash for it from anything you have. The useful part is that the missing page names the missing analysis for you.

Third, a placeholder band wide enough to contain both answers. The Sahyadri new-lab margin band runs from −11.65% to +14.60%. A recommendation that would flip between the two ends of that band is a coin toss with an exhibit attached.

Fourth, a dots test that fails on "what would change it". Authors argue with this one. If nobody reading the dots can say what evidence would overturn the recommendation, the work has produced a position, and a position isn't a finding yet.

What you do next is a communication problem. Take the ghost deck to the sponsor as it stands, on paper, with the gap visible, and ask for either the data or the extra time. Two things make that conversation survivable. The ghost deck shows what has been done, so the request isn't an admission of drift. And the missing page is named, so you're asking for one analysis with a stated cost and not an indefinite extension.

The alternative is the deck that gets built anyway, where the missing page becomes a page with a hedge on it. In the room, the one person who reads carefully finds the hedge. That costs more than the delay, and it's paid in the currency that's hardest to rebuild.

12. The scaffold, and how to use it

You get a ghost-deck scaffold, TPL-ghost-deck, as a fillable storyboard: a page-count header, a numbered card grid, and per-card fields for the five entries. Its filled worked example is the storyboard scaffold itself. The artefact demonstrates its own use, with no synthetic business case behind it.

The header carries four things you fix before drawing: who the audience is and what they're deciding, the slot and discussion reserve in minutes, the computed page ceiling, and the six questions the dots test will score against. Fixing those four takes about ten minutes, and it's the part that makes the rest fast, because every later judgement, cut, order, page count, becomes a comparison against something written down.

The card grid holds one card per intended page, each with the dot, dash, sketch box, source and status fields. Limiting the status field to four values is deliberate. Free text collapses into optimism, and computed, disclosed, estimated, missing are easy to tell apart at a glance when the cards are laid out.

Below the grid, the scaffold reserves three lines that most people forget. The cut list, one line per cut page with the reason. The build plan, in readiness order and not page order. And the stop condition, the trigger that sends you back to the cards.

Use the scaffold on paper first and in the file second, or the other way round, but don't use both at once during the same pass. The version that matters is the one you're willing to reorder. Keeping two copies in step is a job that grows until it's bigger than the deck.

{
  "skill": "Attacking a ghost deck's dot sequence for leaps, duplicates and label headlines before any page is built",
  "use": "After you have written your own dots and dashes by hand, paste the numbered dot list alone, with no dashes and no exhibits, into a chat assistant and make it act as a reader who has never seen the analysis. Ask it to answer your six brief questions from the dots alone, to name every adjacent pair where a reader must supply a missing sentence, and to flag any two dots making the same claim. Use the output as a list of suspects to check by hand, never as a verdict.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "A plain text file holding the dot list, so no slide exists yet", "A spreadsheet or Python for the page-count and rework arithmetic"],
  "prompt": "Below is a numbered list of slide headlines from a storyboard. You have not seen the underlying analysis and you must not assume anything that is not written. 1) Answer these six questions using only the list, and write 'not answerable from the headlines' where you cannot: <QUESTION 1> ... <QUESTION 6>. 2) For every adjacent pair of headlines, say whether a reader must supply a missing sentence to get from one to the next, and if so, write the sentence they must supply. 3) Name any two headlines that make the same claim in different words. 4) Name any headline that states a subject rather than a claim. Do not rewrite my headlines and do not suggest new slides.",
  "verify": "The assistant is scoring readability, not truth, so nothing it returns is a finding until you check it. Every leap it names must be checked against your own cards: either you can point at the dash that closes the gap, in which case the dash belongs in the dot, or you cannot, in which case you have found a missing page and a missing analysis. Every duplicate it names must be checked by comparing the two numbers, since two dots quoting different figures are not duplicates however similar the wording. Every figure in a dot it approves of must still be traced to your own computation, because an assistant reading a headline has no way to know whether 4.6% is the ratio you computed or a number you mistyped. If the assistant answers a question from the dots that your own data cannot support, that is the most valuable output in the run: your headline is overstating.",
  "verify_note": "Assistant behaviour on this task changes between model versions; re-run your own six-question check by hand at least once before trusting the pattern (checked September 2026).",
  "diy": "You must be able to run the read-only-the-dots test cold, on someone else's storyboard, and produce the three failure signatures with the pages named. That is the gated skill, and it is the one the later branches assume: the title-stack work in PR4.02 is this same test applied to a built deck, and the capstone runs it under a time limit on an unseen brief. An author who can only find leaps by asking a machine cannot find them in a meeting, which is where the finding is actually needed.",
  "market": "IN"
}

13. The refusal

The refusal is a rule you apply to yourself: the tool doesn't open until the dots hold.

Held means three things, all checkable. The dots pass the six-question test at five of six or better. Every page has a dash naming an exhibit form, a figure and a source. And the card count is at or under the ceiling that the airtime arithmetic produced.

It sounds like self-control, and it's mostly economics. Everything before the refusal costs about four minutes a page. Everything after it costs between five and fifty-five. Make every judgement you can on the cheap side of the line, and let the refusal stop the expensive side from starting early and eating the judgement.

Opening the tool feels like starting, and sketching feels like preparing to start. Under time pressure people do the thing that feels like starting, so the refusal matters most where it's hardest to apply. Under a deadline, shorten the ghost pass instead of skipping it. Four minutes a page becomes ninety seconds a page. The dots still get written, the sketches get cruder, and the sequence still gets read before anything is built.

There's one exception. A single page, built to answer one question for one person, with no argument to sequence, doesn't need a storyboard. Sketch the exhibit, build it, send it. The method scales down to nothing, and pretending otherwise makes it easy to dismiss.

Common mistakes & how experts think differently

1. Sketching in the tool. Opening the software and typing headlines onto blank slides feels like sketching. It isn't, because the pages aren't cheap and you can already see the fonts. The tell is that you've adjusted something. An expert keeps the sketching medium different from the building medium on purpose, so the two activities can't blur into each other.

1. Writing labels and calling them dots. "Q4 revenue performance" on a card looks like a headline and carries no claim, so the page can't be attacked, can't be cut on its merits and can't be ordered against its neighbours. The habit comes from describing the analysis you did when you should be stating the finding you have. An expert writes the sentence a reader would repeat afterwards, then checks that it contains a number.

1. Sketching one form and building it. The first form that comes to mind is usually the shape of the source table, not the shape of the claim. Three sketches cost 4.5 minutes against 55 minutes for one build, so the cost of checking is about eight percent of the cost of being wrong once. An expert draws the alternative even when confident, mostly because the alternative occasionally wins.

1. Deciding the page count after the pages exist. A count decided at the end is decided by how much material there is, and that's the wrong input. A calendar entry fixes airtime, and it doesn't care how much work you did. An expert computes the ceiling from the slot and the reserve before drawing the first card, and treats it as a hard constraint and never as a target.

1. Filling a placeholder with a plausible number. Write 25% into a gap because it looks reasonable, and you've created a figure the team will start defending within a day. A plausible placeholder anchors harder than an obviously fake one. An expert computes the band the disclosure permits, writes the band, and treats a wide band as a finding about what can be claimed, never as an inconvenience.

1. Reviewing a ghost deck by presenting it. Walking a reviewer through the cards supplies every missing connection out loud, and the deck scores well because the author is the connective tissue. The test is only meaningful when the reader is alone with the dots. An expert hands the list over and takes notes in silence.

1. Treating the cut list as a defeat. Authors delete cut cards because a cut page feels like wasted work. It's the opposite. A page cut at the card stage costs four minutes and saves between 25 and 55, and the cut list is the index that answers "did you consider" without a rebuild. An expert keeps every cut card with its reason line attached.

1. Building in page order. Page order is the reader's order and has nothing to do with which pages are ready, which are blocked and which are most likely to change. Building page 1 first also means building the title page first, which is the least informative work available. An expert sends the data requests, builds the argument's spine, and does the furniture last.

1. Sketching a story that is finished before the analysis is. You can produce a ghost deck for an analysis that hasn't answered its question, and it'll look like a deck. The signals are on the cards: a recommendation page with no dash, a leap no page could fill, a band that contains both answers. An expert reads those signals as the ghost deck working correctly, takes the cards to the sponsor, and asks for the data. They don't build around the hole.

1. Rewriting the dots after the build. Once the pages exist, a dot that no longer matches its exhibit gets edited to match the exhibit, because the exhibit is expensive and the sentence is cheap. That turns the whole method upside down, because the claim should decide the chart. An expert who finds a mismatch after the build treats it as a live question about which one is wrong. About a third of the time, the answer is the chart.

Worked examples

Every figure below comes from one computation, and the Verify section that follows restates each one at full precision. Two synthetic companies carry the examples. Sahyadri Diagnostics Ltd is a synthetic Indian pathology-lab chain, and Cardinal Fleet Services Inc. is a synthetic US commercial-fleet maintenance business. Neither exists, but both sets of accounts tie internally.

Worked example 1: What building first costs (US, $)

A fourteen-page deck is scoped: eight exhibit pages, four text pages and two section dividers. Building in the tool from data that already exists costs 55 minutes for an exhibit page, 25 for a text page and 5 for a divider. A colleague will review for 30 minutes. Experience with this reviewer and this audience says the review will cut five pages, three exhibits, one text page and one divider, and will send exhibits back where the chart form does not carry the claim. Compare building first with sketching first, at 4 minutes for a ghost page.

Solution. Step 1, price the build. 8(55) + 4(25) + 2(5) = 440 + 100 + 10 = 550 minutes for all fourteen. The nine survivors cost 5(55) + 3(25) + 1(5) = 275 + 75 + 5 = 355, so the five cut pages carry 550 − 355 = 195 minutes.

Step 2, run Path A, build first. Build 550, review 30, then rebuild two exhibits at 55 each because the form error is only visible once the chart exists. Total 550 + 30 + 110 = 690 minutes, or 11.50 hours.

Step 3, run Path B, sketch first. Now run it the other way. Ghost 14 pages at 4 minutes each is 56. The same 30-minute review cuts the same five pages, and because a sketch names the chart form, it also catches one of the two form errors. Build the nine survivors, 355, then rebuild the one exhibit the review missed, 55. Total 56 + 30 + 355 + 55 = 496 minutes, or 8.27 hours.

Step 4, verify by an independent route. The saving should equal the wasted work in Path A less the cost of the ghost pass. Path A's waste is the 195 minutes inside the cut pages plus the 55-minute extra rebuild, which is 250 minutes. 250 − 56 = 194, and 690 − 496 = 194. The two routes agree, which is the check that the model is not double-counting the review.

Step 5, find the break-even. The average cut page carried 195 / 5 = 39 minutes of build work, so the ghost pass repays itself at 56 / 39 = 1.44 cut pages.

Step 6, price it in both currencies. At a stated blended rate of $95 an hour, Path A costs $1,092.50 and Path B $785.33, a saving of $307.17. At ₹2,400 an hour the same deck costs ₹27,600 against ₹19,840, saving ₹7,760. Swap in your own rates, since these are assumptions. The 28.1% saving is the part that travels.

Read it. The 194 minutes is nice to have. The number that wins the argument is 1.44 pages. Nobody will adopt a method that needs a fifty-percent cut rate to pay for itself. This one pays at a cut rate any reviewer doing the job will exceed, so the argument for it survives contact with a sceptical team lead. And notice what the model doesn't claim. It doesn't say sketching makes building faster. Build time per surviving page is identical on both paths, at 39.4 minutes. The saving is entirely in which pages get built.

Worked example 2: Eleven pages down to nine (India, ₹)

A teardown of Sahyadri Diagnostics is finished. The audience is a board sub-committee reviewing capital allocation, with a 60-minute slot of which 35 minutes are reserved for discussion. The analyst's first-pass storyboard runs to eleven cards. Convert it into a ghost deck that passes the dots test.

The analysis behind it, all figures in ₹ crore:

NOPAT is EBIT after a 25.17% tax rate, being the 22% base rate with surcharge and cess (confirm the current rate at incometax.gov.in before reusing it; checked September 2026). The analyst's own weighted average cost of capital estimate is 12.5%.

Solution. Step 1, set the ceiling before reading the cards. Airtime is 60 − 35 = 25 minutes. A committee page needs about 2.5 minutes, so the ceiling is 25 / 2.5 = 10 pages. Eleven cards is already over.

Step 2, run the dots test on the first pass. The six questions agreed with the sponsor: what is the recommendation, what is the evidence, what would change it, what does it cost, who owns the first step, when is it re-tested. A colleague reading only the eleven headlines answers three: the recommendation, the cost and the re-test date. Score 3 of 6, 50.0%, against a pass mark of 5 of 6, 83.3%.

Step 3, name the defects. Four cards are label headlines: "Diagnostics market context and Sahyadri's position", "Revenue and lab-count trends, FY22 to FY26", "Invested capital and NOPAT bridge" and "Catchment survey methodology and sample". One card runs to sixteen words. And the card carrying the actual finding is at position six, behind five pages of build-up. Notice that none of these defects is a wrong number.

Step 4, compute the numbers the dots will carry. Revenue CAGR over four years is (731.0/412.0)^(1/4) − 1 = 15.4131%. Lab count compounds at (198/96)^(1/4) − 1 = 19.8391%, and revenue per lab at −3.6933%, a total fall of 13.9747% over the window. The identity holds: 1.1541308 = 1.1983909 × 0.9630671. Incremental return on the new capital is ΔNOPAT / ΔIC = 22.6742 / 493.0 = 4.5992%, which is 7.9008 points below the 12.5% hurdle. Splitting FY26 by cohort, the 142 labs open at the end of FY24 billed ₹645 crore and the 56 opened since billed ₹86 crore, so revenue per mature lab is ₹4.5423 crore against ₹1.5357 crore for a new one, a ratio of 33.81%. Same-lab growth for the mature cohort is 645/601 − 1 = 7.3211% against total growth of 10.4230%. Clearing the hurdle on the new capital needed 493.0 × 0.125 = ₹61.625 crore of extra NOPAT against the ₹22.6742 crore delivered, a shortfall of ₹38.9508 crore. At the FY26 NOPAT margin of 10.7977%, that implies ₹360.7309 crore of extra revenue, 49.35% of FY26 sales. Even billing every one of the 198 labs at the mature rate produces ₹899.3662 crore, a gain of ₹168.3662 crore, which closes 46.67% of that gap.

Step 5, rewrite and reorder. Six edits. Move the finding to page 1. Rewrite two label cards as claims. Tighten the sixteen-word card. Cut two pages. Nine cards survive, in this order.

Nine dots, mean 11.2 words, longest 13, none over the cap.

Step 6, record the cuts and price them. The market-context page answered none of the six questions and went. The methodology page answered "what would change it" less well than an appendix note making the same point, so the claim moved and the page went. Those two pages would have cost 55 + 25 = 80 minutes to build. As cards they cost 2 × 4 = 8. That's 10 to 1, on two pages of one deck.

Step 7, re-score. The rewritten nine answer all six questions. The recommendation is on page 9 and the cost on page 6. The evidence runs across the first seven pages. What would change the call is on page 8. Page 9's dot already carried the re-test date, and the owner of the first step went into its dash. That fixed it at dash level, with no new page. Score 6 of 6, 100.0%.

Read it. The largest gain came from moving one card. The finding was there in the first pass, correctly computed, and nobody could see it because it sat behind five pages of context. That's the ordinary condition of a first draft. It isn't a writing problem. The analyst built the deck in the order the work happened, when the reader needs a different order. What the exercise doesn't show is whether 4.6% is the right number. The ghost deck tests whether the argument lands, and a wrong number lands just as well as a right one. So every dash names a source, and every figure gets recomputed before the build.

Worked example 3: One card, written properly (India, ₹)

Take the cohort card from the Sahyadri deck. The analyst's first version reads:

Dot. "Labs opened in the last two years bill a third of what a mature lab bills." Dash. "Revenue per lab has fallen from ₹4.29 crore to ₹3.69 crore, and the new cohort is much less productive than the mature one, which explains most of the ROIC decline."

Repair both, and state what the repair reveals.

Solution. Step 1, count the dot. Sixteen words against a fourteen-word cap. Tighten by replacing the relative time expression with a date, which also survives a reader meeting the page in six months: "Labs opened since FY24 bill a third of a mature lab's revenue." Twelve words, same claim, same number.

Step 2, check the number the dot claims. 1.535714 / 4.542254 = 33.8095%, which rounds to a third. The claim is defensible at the stated precision. Recompute the two inputs: 86.0 / 56 = 1.535714 and 645.0 / 142 = 4.542254. Cross-check that the cohorts tie: 645.0 + 86.0 = 731.0, the FY26 total, and the weighted cohort figure (142 × 4.542254 + 56 × 1.535714) / 198 = 3.691919 reproduces the blended revenue per lab.

Step 3, read the dash again. It makes three claims, not one. Revenue per lab fell from 4.29 to 3.69. The new cohort is less productive. That explains most of the ROIC decline. Only the second is the dot's evidence. The first belongs to page 2, which already carries it. The third is a causal claim that has not been tested anywhere in the analysis.

Step 4, write the dash properly. "Bar pair, FY26 revenue per lab by cohort: mature 142 labs at ₹4.542 crore, opened-since-FY24 56 labs at ₹1.536 crore, ratio 33.8%. Source: internal cohort split, ties to FY26 revenue of ₹731 crore. Status: computed."

Step 5, deal with the orphaned third claim. The causal claim, that the cohort mix explains most of the ROIC decline, is either a page or it's nothing. Written as a dot it would read "Cohort mix explains most of the ROIC fall from 14.6% to 9.0%". The dash for it doesn't exist, because no analysis has decomposed the 5.6223-point ROIC fall into mix and margin. So the card gets drawn with status missing, and it becomes the third item on the data-request list.

Read it. A three-sentence dash is nearly always three claims wearing one sentence's clothing, and pulling them apart is the cheapest audit you can run on a finished analysis. Here it produced one clean page, one duplicate removed and one real hole discovered, in about four minutes. Notice what the tightened dot gave up, too. "In the last two years" was friendlier, and "since FY24" is checkable. Where the two conflict, take checkable.

Worked example 4: Three sketches in the time of one build (US, $)

Cardinal Fleet Services has grown revenue from $148.0m in FY22 to $219.9m in FY26 while vehicles serviced went from 412,000 to 441,000. The card's dot is "Price drove 85.5% of Cardinal's $71.9m revenue increase, volume 14.5%". Sketch three candidate exhibit forms in about ninety seconds each and choose one.

Solution. Step 1, compute what the dot claims. Revenue per vehicle is 148,000,000 / 412,000 = $359.2233 in FY22 and 219,900,000 / 441,000 = $498.6395 in FY26. Hold price at the FY22 level and apply FY26 volume: 441,000 × 359.2233 = $158.4175m. Volume therefore contributes 158.4175 − 148.0 = $10.4175m and price contributes 219.9 − 158.4175 = $61.4825m. The two sum to $71.9m exactly, and the price share is 61.4825 / 71.9 = 85.5112%. The compound rates carry the same story: revenue at 10.4055%, volume at 1.7151%, price at 8.5439%, with 1.104055 = 1.017151 × 1.085439.

Step 2, sketch form A, a dual-axis line. Revenue on the left axis, vehicles on the right, five years across the bottom. It takes ninety seconds, and it fails the standing rule before it fails anything else. Put two series with different units on two axes and the drawer's scaling choice manufactures the apparent relationship, while nothing on the chart tells a reader that a choice was made. Rejected on the rule, and rejected again on the test in step 5.

Step 3, sketch form B, a waterfall bridge. Four bars: FY22 revenue at 148.0, a rising step of 10.4 labelled volume, a rising step of 61.5 labelled price, and FY26 revenue at 219.9. The accent goes on the price step. Ninety seconds, and the drawing already contains the claim.

Step 4, sketch form C, a slope chart. Revenue per vehicle at FY22 and FY26, two points and a line, $359.22 rising to $498.64. It shows the price move cleanly and says nothing about volume, so it can't carry a claim about shares of an increase. Keep the sketch, though. It's the right form for a different dot on the price page.

Step 5, apply the closing test. Write the number the dot claims, 85.5%, on each drawing and ask whether the shape would show it. Form A shows two rising lines and no decomposition at all. Form C shows a price change and no share. Form B's two middle bars are in the ratio 61.5 to 10.4, so the reader sees 85.5% without being told it. Form B is the choice.

Step 6, price the exercise. Three sketches at 1.5 minutes each is 4.5 minutes, against 55 minutes to build one exhibit. For the cost of a single built chart you could draw about 36 candidates. Drawing three is cheap insurance against a rebuild, and here it prevented one. The analyst's instinct was form A, which is the shape the source table is stored in.

Read it. Nobody rejected the dual-axis chart for being ugly. It went because the drawing can't display the number the sentence claims, and you can put any sketch to that test in five seconds. Hold on to the discarded slope chart. Sketches are cheap enough to keep, and form C comes back as the right exhibit two pages later.

Worked example 5: Five pages for a client read-out (US, $)

A commercial due diligence has found that Cardinal's revenue growth is mostly price, and that most of the price growth lies outside the contractual escalator. The master service agreements cap annual increases at 4.5%. The buyer's model assumes 9.5% revenue growth. The read-out has a 45-minute slot with 25 minutes reserved for the client's questions. Storyboard it.

Solution. Step 1, set the ceiling. Airtime is 45 − 25 = 20 minutes at 2.5 minutes a page, so eight pages. The finding needs five. Coming in under a ceiling is allowed. Coming in over it isn't.

Step 2, compute the evidence. Realised price growth is 8.5439% a year against the 4.5% cap. The residual is multiplicative, 1.085439 / 1.045 − 1 = 3.8698%. Subtraction gives 4.0439 points, so the additive shorthand overstates the residual by 0.1741 points. Now run the counterfactual in levels. Price compounding at exactly 4.5% from $359.2233 reaches $428.3805 in FY26 against the realised $498.6395, and at 441,000 vehicles that gap is (498.6395 − 428.3805) × 441,000 = $30.9842m, which is 14.0901% of FY26 revenue. The top five accounts carry $118.7m, 53.9791% of revenue. Three of the five, carrying $71.4m or 32.4693% of revenue, renew inside eighteen months. Those three accounts alone cover the uncontracted billing 2.30 times over.

Step 3, price the consequence for the model. If out-of-scope billing reverts to the contractual escalator, revenue grows at 1.045 × 1.017151 − 1 = 6.2923% against the model's 9.5%, a gap of 3.2077 points. Compounded over five years from $219.9m, that's $346.175m against $298.355m, a difference of $47.820m, or 13.8138% of the buyer's year-five revenue. Take a 14% EBITDA margin and a 9.0x exit multiple. Both are deal assumptions stated up front, and neither is an observed figure. On those, the difference is $6.695m of exit EBITDA and $60.253m of enterprise value.

Step 4, score the first pass. The analyst's five cards open with "Revenue and volume growth, FY22 to FY26", a label, and close with the finding. A reader given only those five headlines answers four of the six questions agreed for a diligence read-out: what the finding is, what the evidence is, what it does to the model, and what the ask is. Score 4 of 6, 66.7%, below the pass mark. Four of six sounds respectable until you remember where the pass mark sits.

Step 5, repair. Move the finding to page 1. Rewrite the label as a claim. Add the two renewal dates to page 5's dash, which answers what would change the finding and when it resolves.

Five dots, mean 11.2 words, longest 12. Score after the repair, 6 of 6.

Step 6, note the precision the dot gives up. Page 3's dot says "4.5 of the 8.5 points", which is the additive framing and is 0.1741 points generous to the residual. That's acceptable in a fourteen-word sentence and unacceptable in the dash, which carries the multiplicative figure of 3.8698% and the identity 1.045 × 1.038698 = 1.085439. A dot may round. A dash may not.

Read it. Every page here is a claim about the same finding seen from a different angle, and a five-page read-out should look like that. Nobody expected to need page 4. It exists because the client's first question will be whether the uncontracted billing is diversified, and it isn't. Three accounts carry it, and they renew inside the diligence window. The ghost deck can't tell you whether those renewals will hold, so the dash carries a date and no forecast.

Worked example 6: The number nobody disclosed (India, ₹)

Page 5 of the Sahyadri deck was going to claim that the new-lab cohort is loss-making at the EBITDA line. Sahyadri doesn't disclose EBITDA by cohort. Decide what the page can fairly claim.

Solution. Step 1, state what is known. FY26 EBITDA is ₹183.481 crore, being 731.0 × 0.251. The cohort revenue split is computable: ₹645 crore mature, ₹86 crore new, tying to ₹731 crore.

Step 2, bound the unknown. Assume a mature-lab EBITDA margin and take the residual as the new cohort's.

Step 3, read the band. Across a plausible range of mature-lab margins the new cohort's margin runs from −11.65% to +14.60%, a band 26.25 points wide. A band that contains both a healthy number and a loss cannot support a claim in either direction.

Step 4, rewrite the page. The dot that was going to say "the new cohort is loss-making" becomes a card with status missing and a different job. Two options were live. One was to rewrite the claim to something the data supports, which is the revenue-per-lab ratio of 33.81% already carrying page 3. Then the page is a duplicate and gets cut. The other was to keep the page and make the band itself the claim: "Disclosure cannot say whether new labs make money; the band is 26 points wide." That's fourteen words, right at the cap. It's a claim, and a reader can disagree with it.

Step 5, price the data request. One question to the finance team, the mature-cohort EBITDA margin, collapses the band to a point. Nothing else on the request list has that property, so it goes first. The page gets built when the answer lands, or rewritten if it doesn't.

Read it. The band measures what the disclosure can carry, and it took about four minutes to establish. Now look at the trap you just avoided. Write 25% into the gap because it falls between the extremes, and you'd have a page, a chart and a claim. A team that had forgotten where the number came from would then defend all three for a fortnight. A figure that looks reasonable anchors harder than one that looks invented.

Worked example 7: The page count decided by a calendar entry (US, $)

A deal team has a 24-page deck ready and four possible audiences. Compute the page ceiling for each and diagnose what happens when the deck is taken to the wrong one.

Solution. Step 1, compute airtime and the ceiling.

Same deck, four rooms, four different ceilings.

Step 2, read the reserve column before the page column. The committee reserves 55.6% of its slot for argument and the working session 33.3%. That difference describes what each audience is for. A committee that has read the pack beforehand is buying an interrogation, so its pages exist to be attacked, and showing them isn't the point.

Step 3, price a pre-read separately. A pre-read has no airtime. Price it as a reading budget instead. At 12 minutes of real attention and about 0.75 minutes to read a page carrying a full dash line, the ceiling is 16 pages, twice what the same audience can absorb in the room.

Step 4, diagnose the 24-page deck in the investment committee. Twenty minutes of airtime across 24 pages is 20 / 24 = 0.833 minutes a page, against the 2.5 minutes the format needs. That's a three-fold over-build. You might expect the presenter to run out of time at the end. What actually happens comes earlier. The committee stops the presenter on page 4 to ask the question page 19 answers, and the remaining sixteen minutes go on navigating.

Step 5, state the repair. Don't shorten the deck by deleting pages at random. Set the ceiling at eight, rank the 24 cards against the six questions, keep the eight that answer them, and move the rest to an indexed appendix that the pre-read points at. The pre-read ceiling of 16 tells you how much of that appendix can sit in the sent version.

Read it. Page count is arithmetic, done from a calendar entry that exists before any work is scoped. An author who computes the ceiling on day one builds a different deck from one who discovers it on the morning of the meeting. The difference is which pages exist, and effort has nothing to do with it. The model leaves out a presenter who talks faster, and that's deliberate. Twenty-four pages delivered in 20 minutes produces the same meeting with the argument taken out of it.

Verify

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

Rework model. Build 8(55) + 4(25) + 2(5) = 550.000000; survivors 5(55) + 3(25) + 1(5) = 355.000000; cut pages 195.000000. Ghost 14 × 4 = 56.000000. Path A 550 + 30 + 110 = 690.000000 minutes, 11.500000 hours. Path B 56 + 30 + 355 + 55 = 496.000000 minutes, 8.266667 hours. Saving 194.000000 minutes, 3.233333 hours, 28.115942% of Path A. Waste 195 + 55 = 250.000000 minutes, 36.231884% of Path A; identity 250 − 56 = 194.000000. Average cut page 195/5 = 39.000000 minutes; break-even 56/39 = 1.435897 pages; net at 1, 2, 3 and 5 cut pages −17.000000, +22.000000, +61.000000, +139.000000. Costs $1,092.500000 and $785.333333, saving $307.166667; ₹27,600.000000 and ₹19,840.000000, saving ₹7,760.000000. Build per surviving page 355/9 = 39.444444 minutes on both paths.

Sahyadri. EBITDA 412.0 × 0.265 = 109.180000, 496.0 × 0.278 = 137.888000, 585.0 × 0.284 = 166.140000, 662.0 × 0.269 = 178.078000, 731.0 × 0.251 = 183.481000. EBIT 75.180000, 95.888000, 114.140000, 114.078000, 105.481000. NOPAT at (1 − 0.2517): 56.257194, 71.752990, 85.410962, 85.364567, 78.931432. ROIC 14.612258%, 15.266594%, 14.427527%, 11.677779%, 8.989913%; fall 5.622346 points. Revenue per lab 4.291667, 4.203390, 4.119718, 3.871345, 3.691919. Revenue CAGR (731.0/412.0)^0.25 − 1 = 15.413082%; labs (198/96)^0.25 − 1 = 19.839086%; revenue per lab −3.693289%, total fall −13.974698%; identity 1.154130824 = 1.198390860 × 0.963067110. Labs added 102. ΔNOPAT 22.674238, ΔIC 493.000000, incremental ROIC 4.599237%, gap to 12.5% −7.900763 points. Cohorts: mature 142 labs at 645.0, new 56 at 86.0, sum 731.000000; FY25 601.0 + 61.0 = 662.000000. Revenue per lab 645.0/142 = 4.542254 and 86.0/56 = 1.535714; ratio 33.809524%; weighted 3.691919 reproduces the blend. Same-lab growth 645.0/601.0 − 1 = 7.321131%; total growth 10.422961%. Required NOPAT uplift 493.0 × 0.125 = 61.625000; shortfall 38.950762; FY26 NOPAT margin 10.797734%; implied extra revenue 360.730902, 49.347593% of FY26 revenue. All labs at the mature rate 198 × 4.542254 = 899.366197; gain 168.366197; closes 46.673627%. Screen ₹3.2 crore is 2.083721x the new-lab average and 70.449612% of the mature rate; 56 labs at the screen adds 93.200000 crore of revenue and 10.063488 crore of NOPAT, lifting incremental ROIC to 6.640512%.

Placeholder band. FY26 EBITDA 183.481000. At 26.5%: mature 170.925000, residual 12.556000, margin 14.600000%. At 28.0%: mature 180.600000, residual 2.881000, margin 3.350000%. At 30.0%: mature 193.500000, residual −10.019000, margin −11.650000%. Band width 26.250000 points.

Cards and cuts. First pass 11 pages, ceiling 10, kept 9, cut 2. Cut pages' build cost 55 + 25 = 80.000000 minutes against 2 × 4 = 8.000000 minutes as cards, ratio 10.000000. India dots: 9, mean 11.222222 words, longest 13. Failing dot 16 words, tightened 12, cap 14. Label headline 26 words to dot 11 words, ratio 2.363636. Dots test: Sahyadri 3/6 = 50.000000% then 6/6 = 100.000000%; Cardinal 4/6 = 66.666667% then 6/6 = 100.000000%; pass mark 5/6 = 83.333333%.

Cardinal. Revenue per vehicle, the two years the example states: FY22 148.0/0.412 = 359.223301, FY26 219.9/0.441 = 498.639456. CAGRs: revenue 10.405543%, volume 1.715080%, price 8.543928%; identity 1.104055431 = 1.017150800 × 1.085439288. Increase 71.900000; FY26 volume at FY22 price 158.417476; volume 10.417476, price 61.482524, sum 71.900000; shares 85.511160% and 14.488840%. Escalator 4.5%; residual multiplicative 3.869788%; additive 4.043928 points; overstatement 0.174140 points. Counterfactual price at the 4.5% cap 359.223301, 375.388350, 392.280825, 409.933462, 428.380468; FY26 gap 70.258988 per vehicle, 30.984214 $m, 14.090138% of revenue. Top five 118.700000, 53.979081%; three accounts 71.400000, 32.469304% of revenue and 60.151643% of the block, covering the uncontracted billing 2.304399 times. Growth 9.5% against 1.045 × 1.017150800 − 1 = 6.292258%, gap 3.207742 points; year five 346.175099 against 298.355054, difference 47.820046, 13.813832%; EBITDA difference 6.694806; enterprise value difference 60.253257. US dots: 5, mean 11.200000 words, longest 12.

Page counts. Airtime 60, 20, 25, 20; ceilings 20.000000, 8.000000, 10.000000, 13.333333; reserve shares 33.333333%, 55.555556%, 58.333333%, 33.333333%. Pre-read 12 / 0.75 = 16.000000 pages. Over-build 20/24 = 0.833333 minutes a page against 2.5, factor 3.000000. Sketching 3 × 1.5 = 4.500000 minutes against 55 to build, 36.666667 sketches for one build.

Practice set

Work each problem before reading its solution. You'll want a calculator. Numeric answers within 2% of the stated figure count as correct unless a problem says otherwise. Where a problem asks for a sentence, the test is whether a reader could disagree with it and whether it fits inside fourteen words.

P1 (guided). Price the ghost pass (India, ₹). A twenty-page deck is scoped. An exhibit page costs 50 minutes to build and a text page 20. The reviewer is expected to cut six pages, four exhibits and two text pages. A ghost page costs 3 minutes. Compute the build work inside the cut pages, the cost of the ghost pass, the net saving in minutes and hours, and its value at a blended rate of ₹2,400 an hour.

Solution. The cut pages carry 4(50) + 2(20) = 200 + 40 = 240 minutes of build work. The ghost pass costs 20 × 3 = 60 minutes. Net saving 240 − 60 = 180 minutes, which is 180/60 = 3.0 hours, worth 3.0 × 2,400 = ₹7,200. Notice what the calculation ignores, and why that makes it conservative. It gives the ghost pass no credit for the form errors a sketch catches before a chart exists, which in the first worked example was one further rebuild.

P2 (guided). Find the break-even (US). An eighteen-page ghost pass costs 4 minutes a page. The average page this reviewer cuts carries 42 minutes of build work. How many pages must be cut before the ghost pass pays for itself, and what do you tell a team lead who says the reviewer usually cuts one page?

Solution. The ghost pass costs 18 × 4 = 72 minutes. Break-even is 72 / 42 = 1.714 pages, so two cut pages repay it and one doesn't. Tell the team lead straight that at one cut page the pass loses 72 − 42 = 30 minutes. Then point out that a reviewer who cuts one page out of eighteen is either reviewing a very good deck or not reviewing it. The second possibility is the one worth checking, because a review that cuts nothing is usually a review its author walked through.

P3 (guided). Set the ceiling (US). A read-out has a 65-minute slot. The client is expected to spend 30 minutes asking questions. Pages of this kind need 2.5 minutes each. State the ceiling and what you would do with a storyboard that came in at nineteen cards.

Solution. Airtime is 65 − 30 = 35 minutes, so the ceiling is 35 / 2.5 = 14 pages. Nineteen cards is five over. Rank all nineteen against the six questions the brief will be scored on, keep the fourteen that answer them or directly support a page that does, and move the other five to an indexed appendix with the reason recorded on each card. Cutting by ranking works and cutting by asking "is this page good" doesn't, because every page passes that question when its own author asks it.

P4. Compress a label into a dot (India, ₹). An analyst's card reads: "Analysis of revenue and lab-count trends across the FY22 to FY26 period including a discussion of unit economics and the drivers of the recent margin compression." Count it, write a dot that replaces it using the Sahyadri figures, and state the compression ratio.

Solution. The original runs to 26 words and makes no claim. The dot: "Revenue grew 15.4% a year while revenue per lab fell 14%", which is 11 words, and the ratio is 26/11 = 2.364. The replacement is a claim, because a reader could argue that a falling revenue per lab is the expected arithmetic of opening labs quickly and says nothing about quality. That's a real objection, and the deck answers it two pages later. Nobody can argue with a label headline. It feels safe for that reason, and it's useless for the same one.

P5. Audit five dots. Which of these pass all four dot tests, and name the specific failure for each that does not.

(i) "Q4 channel performance and outlook." (ii) "Marketplace revenue fell from 45.0% to 32.9% of the total, so April's spend plan moves." (iii) "Growth was strong across most channels." (iv) "Quick commerce grew 169.7% while marketplaces shrank." (v) "Marketplaces are now a third of revenue, down from nearly half."

Solution. (ii) and (iv) pass. (i) is a label: no verb doing work, no claim, no number. (iii) carries no number and cannot be disagreed with, since "strong" and "most" survive any outcome. (v) is a duplicate of (ii) making the same claim with rounder numbers, and one of the two goes. Keep (ii), because it names the consequence and so justifies its page. As for counting, (ii) is 15 words and has to lose at least one. The cheapest cut is "of the total", which the exhibit will show anyway.

P6. Choose the sketch (India, ₹). A synthetic auto-components maker grew revenue from ₹640 crore to ₹902 crore while units shipped went from 5.10 lakh to 5.62 lakh. The dot is to claim that price, not volume, drove the increase. Compute the split, then choose between a dual-axis line of revenue and units, a waterfall bridge, and a slope chart of realised price.

Solution. Realised price is 640/5.10 = 125.4902 and 902/5.62 = 160.4982 ₹ crore per lakh units. Hold price at the base level and apply the later volume: 5.62 × 125.4902 = 705.2549. Volume contributes 705.2549 − 640 = 65.2549 and price contributes 902 − 705.2549 = 196.7451. The two sum to the 262.0000 crore increase. Price is 196.7451/262.0 = 75.0935% of it. The waterfall is the choice, because its two middle bars stand in the ratio the dot claims and a reader sees 75% without being told. The dual-axis line is rejected on the standing rule and, more usefully, on the closing test. Write 75.1% on the drawing and the shape can't show it. The slope chart shows the price move cleanly and says nothing about volume's share, so keep the sketch for the price page and don't use it here.

P7. Bound a placeholder (US, $). A company reports total EBITDA of $96m. Its mature sites bill $318m of revenue and its new sites $122m. The mature-site EBITDA margin is not disclosed and is thought to lie between 24% and 28%. State the band the new-site margin must lie in, its width, and what the page can claim.

Solution. At a 24% mature margin, mature EBITDA is 318 × 0.24 = 76.32, leaving 96 − 76.32 = 19.68 for the new sites, a margin of 19.68/122 = 16.1311%. At 28%, mature EBITDA is 89.04, leaving 6.96, a margin of 5.7049%. The band runs from 5.70% to 16.13%, a width of 10.4262 points. Unlike the Sahyadri case, the band stays positive throughout. So the page can fairly claim that new sites are profitable at the EBITDA line, and it can't claim how profitable. Write the dot to the claim the band supports: "New sites are EBITDA-positive, between 5.7% and 16.1%, on disclosed splits." Then put the one data request that collapses the band at the top of the list.

P8. Diagnose the sequence. A six-card storyboard reads: (1) Market context and our position. (2) Revenue grew 22% while gross margin fell 3 points. (3) Discounting rose in two of five regions. (4) Discounting rose in the regions where a new entrant launched. (5) Contribution per order fell 11%. (6) Hold list price and fund a regional response. Name every defect and state the repair.

Solution. Three defects. Card 1 is a label answering none of the brief's questions, and it goes. Cards 3 and 4 are near-duplicates, both claiming that discounting rose in specific regions. Keep card 4, which carries the cause, and demote card 3's exhibit to the appendix. And there's a leap between cards 5 and 6. Nothing says why an 11% fall in contribution per order implies holding list price instead of cutting it. So a card is missing and its dash doesn't exist yet, which means the hole is in the analysis itself and not in the storyboard. After the repair the sequence is four cards plus one to be built, which is a better deck and a worse afternoon.

P9. Read a dash (US, $). A card's dash reads: "Bar pair, incremental ROIC against hurdle: ΔNOPAT $31.4m over ΔIC $402m, hurdle 9.0%. Source: FY22 and FY26 balance sheets. Status: computed." Compute the figure and write the dot.

Solution. 31.4 / 402 = 7.8109%, which is 7.8109 − 9.0 = −1.1891 points below the hurdle. A dot: "New capital earned 7.8% against a 9.0% hurdle over four years", which is 11 words and carries both numbers. Two things to notice. The dash names its source and its status, so a reader can attack the claim without waiting for the chart. And the gap is small enough that the precision matters. At a 7.5% hurdle the same analysis reverses, so the hurdle's own derivation belongs in the appendix and the dash should point at it.

P10. Cut to a ceiling. A twelve-card storyboard faces an eight-page ceiling. Four cards must go. State the procedure and the two things you record for each cut.

Solution. Rank all twelve against the six questions agreed with the sponsor. Every card either answers one of the six, directly supports a card that does, or does neither. Cut from the third group first, in order, until eight remain. If the third group is smaller than four, cut the weakest supporters, never the answers. For each cut record two things: which of the six questions the card failed to answer, and where its claim now lives, whether that is an appendix page, another card's dash or nowhere. Keep the cut cards. When someone asks on Thursday why the competitive-landscape page is gone, "it answered none of the six questions you asked for" is a different conversation from "it felt weak", and the card is still on the desk if the answer proves wrong.

P11. Order the build. Nine validated cards. Pages 1, 2, 3, 4, 5, 7 and 9 are computed. Pages 6 and 8 are estimated, each waiting on one data request. Pages 1, 5 and 9 carry the recommendation and its central evidence. Write the build order and the stop condition.

Solution. Send the two data requests before building anything, so they run while other work proceeds. Then build 1, 5 and 9, the argument's spine, because those are the pages most likely to change shape once they're real and you want that discovery early. Then 2, 3 and 4, which are computed and have no dependencies. Then 6 and 8 when their data lands, or rewrite their dots to whatever the bound supports if it doesn't. Then page 7, the dividers, the source lines and the appendix index, last, because furniture changes with the sequence and is the most attractive procrastination available. The stop condition, written at the top: if either data request comes back outside the bounded band, pages 5 and 9 change and the whole sequence is re-read before any further building.

P12. Refuse, or proceed. A storyboard scores 5 of 6 on the dots test. Every page has a dash except the recommendation page, whose dash reads "as discussed with the team". The card count is nine against a ceiling of ten. The meeting is in two days. Decide.

Solution. Refuse. The score is what makes the decision look harder than it is. Five of six passes the dots test and the count is under the ceiling, so two of the three conditions hold. The third doesn't. A recommendation page whose evidence line is "as discussed" has no evidence, and the missing question in the 5 of 6 is almost certainly "what is the evidence for the recommendation", the one you can't ship without. Two days is enough to find the dash or to change the recommendation to one the analysis supports. It isn't enough to do either after building nine pages. Take the cards to the sponsor today, name the missing analysis, and ask for the data or the smaller claim.

Applied mini-project

Produce a complete ghost deck for the supplied brief, on paper or in a plain text file, and deliver it as five artefacts. Budget about two hours. Don't build anything in a presentation tool. A submission containing a slide file scores zero on the artefact rows, however good it is.

The brief is Cardinal Fleet Services' own board, meeting to decide whether to accept an approach from a financial buyer. The board has seen none of the diligence work. The slot is 45 minutes with 20 minutes reserved for discussion. You have the FY22 to FY26 data used in Worked examples 4 and 5, the escalator cap of 4.5%, the account concentration, and the finding that most price growth lies outside the contractual escalator. The decision is whether to engage or to decline and fix the pricing exposure first.

Notice what changed and what didn't. The analysis is identical to the buyer's read-out. The audience, the decision and therefore the storyboard aren't, and producing a different ghost deck from the same numbers is the point of the exercise.

The first artefact is the header, four lines fixed before any card is drawn. Who the board is and what it is deciding. The slot and the reserve. The computed page ceiling, (45 − 20) / 2.5 = 10 pages. And the six questions the dots test will score against. Write those yourself, for this board. Don't borrow them from the read-out, because a board deciding whether to sell asks different questions from a buyer testing a model.

The second artefact is the card set, one card per intended page, each with all five fields filled: dot, dash, sketch, source, status. Dots at fourteen words or fewer, each carrying its number. Dashes naming an exhibit form, a figure and a source. Sketches drawn badly on purpose, with axes labelled, sort order shown and the accent series marked. At least one card must carry a status other than computed, and if your reading of the data produces none, say so and defend it.

The third artefact is the dots test: your numbered dot list, given to a second reader with the dashes and sketches covered, with their answers to your six questions and your score out of six. If no second reader is available, run it on yourself after a twelve-hour gap and say that you did. Then the repaired list, with each edit labelled as a rewrite, a reorder, an insertion or a cut.

The fourth artefact is the cut list, carrying every cut card, its dot, the question it failed to answer, and where its claim now lives. An empty cut list means one of two things. Either your first pass came in right at the ceiling, which happens rarely and should be defended in one line. Or the storyboard was never over-drawn, which means you did the cutting in your head before the cards existed and skipped the exercise.

The fifth artefact is the build plan with its stop condition. Order it by readiness and argument load. Page number doesn't come into it. List the data requests first, and add one line naming the trigger that sends you back to the cards.

Rubric. Twenty points, passing at 16 with no zero on any row.

Any artefact whose test fails scores zero on its row regardless of how well the submission reads. A beautiful ghost deck whose dots don't answer the board's questions has failed at the only thing it exists to do.

One optional second pass is worth the hour. Redraw the same brief for a 20-minute board slot with 10 minutes reserved, giving a ceiling of 10 / 2.5 = 4 pages. Four cards from the same analysis aren't a shortened version of nine. They make a different argument, and finding out which four is the fastest way to discover which of your nine actually mattered.

Reading & resources

Core (do these).

• Barbara Minto, The Pyramid Principle: the structural source for why a claim goes first and how the levels beneath it must answer the question the level above provokes. Read the material on the introduction structures and on ordering before the deduction chapters. It's the book behind the reorder that does most of the work in Worked example 2. [Paid] [Intermediate]
• Gene Zelazny, Say It With Charts: the original message-to-form mapping, written for people who have to choose a chart in ninety seconds. The compact grammar table you used for sketching descends from it. [Paid] [Beginner]
• Cliff Atkinson, Beyond Bullet Points: the closest published treatment of storyboarding before building, including the habit of writing every headline as a sentence. Take the storyboard method and leave the three-act framing, which suits a keynote better than a committee page. [Paid] [Intermediate]
• Cole Nussbaumer Knaflic, Storytelling with Data: the best single volume on the exhibit craft that the sketch stage deliberately defers, and its exercise book, Let's Practice, is where the reps are. [Paid] [Beginner]

Deeper (do these next).

• Dan Roam, The Back of the Napkin: the case for drawing badly on purpose, and a usable taxonomy of what a rough picture can show that a sentence cannot. The part you want argues that a sketch is a thinking instrument. It isn't a draft of a graphic. [Paid] [Beginner]
• Jean-luc Doumont, Trees, Maps and Theorems: dense, short, and the most rigorous published statement of message-first structure for technical audiences. It is the book that makes the fourteen-word cap look generous. [Paid] [Advanced]
• Nancy Duarte, Resonate: on shaping a sequence for an audience you need to move as well as inform. Read it against the honesty constraint the track imposes, because a storyboard built for persuasion needs the same evidence line as one built for a committee. [Paid] [Intermediate]
• Stephen Few, Show Me the Numbers: the reference for what a given form can and cannot display, which is the question the closing sketch test asks. [Paid] [Intermediate]
• Alberto Cairo, The Truthful Art: on the difference between a chart that is defensible and a chart that is merely clear, and the best treatment of how uncertainty should appear in a graphic. Read it alongside the placeholder rule. [Paid] [Intermediate]
• Edward Tufte, The Visual Display of Quantitative Information: the origin of the data-ink argument that PR2.02 owns in full. Read it after the mechanics, or the aesthetics will arrive ahead of the grammar. [Paid] [Advanced]

Free.

• The Financial Times' Visual Vocabulary, a one-page public chart-choice reference organised by message type, useful printed and pinned next to the sketching pad. Availability and edition change; check before relying on a specific version (checked September 2026). [Free] [Beginner]
• The storytellingwithdata blog and its makeover archive, where before-and-after pairs are worked in public with the reasoning attached. The makeovers are more useful than the posts. [Free] [Beginner]
• The track's own scaffold and drills: the ghost-deck scaffold, the storyline drill family with its parent, child and order checks, and the makeover sets that PR1.02 and PR2.02 feed. The drills are where the twenty-four reps live, and reps are what turn the four tests on a dot into something you apply without counting. [Free] [Beginner]

Teach it back & journal

Teach it back. Find someone who builds decks and doesn't storyboard, and make the case in under five minutes without using the words ghost deck or dot-dash. Give them the arithmetic first: fourteen pages, five cut, 690 minutes against 496, and the break-even at 1.44 cut pages. Then hand them a card and have them write one dot for a piece of work they finished last week, counting the words out loud. Most people produce a label on the first attempt and a claim on the second, and watching that happen to someone else is the fastest way to learn to see it in your own drafts.

Then teach the harder half, the part people nod at and don't adopt. Explain why a page that cost four minutes gets killed on its merits and a page that cost fifty-five gets defended. Then ask them to remember the last time they argued for a slide they privately knew was weak. Don't sell the method as a productivity trick. It's a device for keeping your own judgement available at the moment you need it, and the arithmetic is only the argument that buys you permission to try it.

Finish by handing them the six-question test on a real deck of theirs, reading only the titles, and scoring it. Don't soften the score.

Journal. Take the last deck you built and reconstruct its ghost deck after the fact. Write the dot each page would have carried, count the words, and mark every page whose headline proves to be a label. Then settle three things in writing.

Start with the count. How many of your pages would have survived a ceiling computed from the actual meeting length, and which ones would you have lost?

Then find the leap you were covering with speech, and say whether you could have written a dash for the missing page from what you held at the time. Last, price it both ways: the ghost pass you did not run, against the build minutes now sitting inside the pages nobody discussed.

Write the answers as numbers, and leave impressions out. Then write one sentence about the page you'd still have built anyway, knowing it would be cut, and why. That sentence is usually the truest one, and it names the habit the next deck has to break.