The deck: one message per slide, and the slop anti-lever

Three rules turn the pyramid into pages. Governing thought = title slide + executive summary. The answer travels first in the deck exactly as it did in the summary - a partner who reads only page one leaves with the recommendation. Each support argument = a section. Demand, economics, window: three dividers, three chapters, in the pyramid's order. Each evidence point = one slide with one message, and the action title IS the message: "62% of trial teams hit the seat cap in 3 weeks" is a title; the slide beneath it exists to prove that sentence and nothing else.

The elevator test from b5 becomes a page-flip test here: read the titles in sequence, ignoring every body. If the titles alone carry the storyline, the deck works. If a title merely names a topic - "Retention analysis" - the page is freight without an address. And exhibits earn their place by proving the title; anything on the page that does not prove the title is decoration, and decoration is deleted.

A useful pre-draft habit: count the pyramid before opening any tool. The Cadence pyramid has one governing thought, three arguments, and six evidence points that survived b4's quality gate - so the readout deck wants roughly seven content slides plus dividers, and anything much longer is padding wearing a suit. Sizing from the pyramid also gives you the AI brief for free: seven titles, each with its number, is a checklist a model can draft against and a human can verify in minutes. Decks grow when they are drafted before they are counted.

Slop has a signature, and once you see it you cannot unsee it:

Notice that the b5 measures are, collectively, a slop detector. Quantified support and client anchors are the sharpest blades: a slop deck scores ZERO on anchors, because "Cadence", "seat cap", "procurement", and "$30" simply do not appear - and near zero on quantified support, because the average deck hedges where numbers would commit. The lab below makes this concrete: the fluent full-AI draft scores 8 of 100. Read that draft aloud before you see the score; it genuinely flows.

The fix is a contract, not a boycott: AI proposes structure and first-pass language FROM YOUR FACTS. Feed it the source log and the storyline - "here are the seven titles with their numbers; draft the body bullets for slide 3, each ending in its consequence" - and it is a superb accelerant. Never ask it to "write the deck about enterprise SaaS pricing" cold; that prompt has no access to anything that makes the deck yours. The tools in both experiments below are identical. The only difference is who supplies the facts, the so-whats, and the signature - and that difference is 92 points.

Test one - the "and" test: if the action title needs "and" twice, the page is carrying two messages, and it is two slides. Splitting is free; a muddled page is not. Test two - exhibit honesty: the chart shows the number that proves the title, at the size the argument deserves. Two bars, 94 vs 71, beats a 12-line table every time the message is "converted accounts compound" - the table is more data and less proof.

AI's role at the page level: it is genuinely good at suggesting exhibit FORMS - "what is the cleanest way to show a retention gap between two cohorts?" is a fine prompt, and "two bars" is a fine answer. The numbers inside the exhibit come from week 2-3's logged work, full stop. Form from the model, facts from the log, message from you - the one-message rule holds the three together on every page.

"Decoration is deleted" does not mean the supporting material dies - it moves. The appendix is the one-message rule's pressure valve, and it has its own discipline:

• The 12-line retention table goes to the appendix, referenced from the two-bar slide - present for the analyst who asks, absent from the argument's path.
• Methodology and the source log live there too: when a client asks "where does the 62% come from?", the answer is one page-flip away, which is exactly the confidence b3's discipline was built to buy.
• Nothing enters the main flow from the appendix without earning a title of its own. If a backup page keeps getting pulled forward in rehearsal, it was a message all along - promote it, give it an action title, and count the deck again.

AI is useful here in the least glamorous way: "list every exhibit in this draft that does not directly prove its slide's title" is a fine demotion-candidate prompt. The demotion decision, like every other one this week, stays with you.