Causality Engine brand system 7.3: LLM cheatsheet

Source: https://brandbook.causalityengine.ai. Plain text copy: https://brandbook.causalityengine.ai/llms.txt. This sheet is the whole book, compressed for a model. Read it top to bottom once. It is shorter than it looks.

How to use this sheet

You were probably handed this link and asked to write, design or check something for Causality Engine. Good. Here is the order of authority.

1. What we are allowed to claim beats everything. If a rule below would make a sentence untrue, drop the sentence.
2. Verbatim lines are verbatim. Quote them exactly or do not quote them.
3. Tokens are exact. Use the hex and the pixel values given. Do not round, tint or "improve" them.
4. Unresolved decisions at the bottom are open. Do not present either side as settled.
5. When this sheet and a chapter of the book disagree, the chapter is newer. Say that you found a disagreement.

If you are unsure whether something is allowed, it is not. Ask the person who sent you.

Identity

• Name: Causality Engine. Two words, both capitalised. Never CausalityEngine, never CE in public copy.
• Category: causal attribution.
• Website: https://www.causalityengine.ai
• What the product does, in one breath: it reads every marketing channel in a brand's Google Analytics 4 export and says which ones actually moved revenue, and how sure it is.

The one line

Use this wherever a stranger meets us for the first time. Word for word.

You spend money on marketing. Some of it works. We tell you which parts, and how sure we are.

Positioning anatomy

Every piece of positioning copy answers six questions. Here is the one line, taken apart. The same data draws the diagram in the book's Contents.

• First line, live: Live on causalityengine.ai, word for word. "You spend money on marketing. Some of it works. We tell you which parts, and how sure we are."
• Second line, draft: Built only from live facts. Not on the site yet. "One €99 read of every channel in your GA4 export shows ecommerce owners and in-house marketing leads what each channel actually contributed, before the next budget decision."

When you write new positioning copy, fill all six before you polish any of them. Quote the first line exactly. Do not present the second line as live copy.

Ideal customer profiles

Two buyers, one method (after Fletch, How to build your ICP). The use case is non-negotiable, the competitive alternative and the problem are required for positioning, and everything else is optional and only narrows the segment.

• ICP 1, The owner-operator: Solo, watching cash flow, just added channels. Every euro is their own.
• ICP 2, The marketing lead: Runs a team, reports every month, gets asked which channels work.

Ad spend is open. The source material names two bands that do not agree, and the brand has not chosen. Do not state a minimum.

What we measure

• Every channel, not just the ads. Paid, organic search, organic social and email go into the same model.
• Key events come from the Advertising tab in Google Analytics 4. Every other channel in the same export goes in alongside them.
• Never describe Causality Engine as an ad tool or a paid-media tool. People say ROAS when they mean ads. Ads are part of the answer. They are not the whole of it.

What every read returns

1. The estimate: what the channel contributed.
2. Its confidence interval: how sure we are, stated as a range.
3. Coverage: the share of real orders the estimate explains.
4. Method label: how the number was produced.

A number without its range and its method is not a Causality Engine number.

What it costs

• €99: one read of every channel in your GA4 export, ads and organic alike.
• €299/mo: continuous reads, for brands that want the number every week rather than once.
• Refundable if the first read does not move a budget decision.
• No contract, no annual lock-in, no pixel and no code change.
• Enterprise pilots are quoted separately and do not carry the refund.
• Lead with the price. Buyers find us through how we are sold, before they care how the model works.

Who we write for

The person who owns the marketing budget: ecommerce owner-operators and in-house marketing leads.

Two registers, always paired

Expert wording shows we know the field. Plain wording is how buyers actually ask. Every expert claim gets a plain twin, and we publish both.

The plain column is written the way people type: lowercase, no category words, often no question mark. Keep it that way.

Plain speak toward customers

• Anything a customer reads should be followable by a five-year-old. Short words. One idea per sentence.
• No category jargon unless its plain twin sits next to it.
• On screen and on the site, use numerals. Spelled-out numbers are a voiceover rule only.

What we never say

1. "AI-powered", "machine learning" or "AI platform". The method is a proprietary causal-inference model.
2. Anything that frames us as an ads-only tool.
3. A statistic, testimonial, rating, customer count or success rate we cannot trace to a named, consenting customer or a dated public source.
4. A competitor's price without the link to their pricing page and the date it was read.
5. A customer name or result that is not already published at https://www.causalityengine.ai/llms.txt.

Names and handles

Warning about X. The X handle uses a capital I where "Causality" has a lowercase l: @Causa + capital I + ityEngine. They look identical in most fonts. Copy it from this table. Never retype it, and never "correct" it.

How we sound

Short sentence. Then a longer one that develops the thought, turns on a clause and lands somewhere the reader did not expect. Then short again. That rhythm is the voice. Uniform sentence length is the clearest sign a draft is not ours.

• Periods, never exclamation marks. Authority does not shout.
• Exact numbers. €4,314, not "about €4,000". A round number reads as an estimate. When a figure is illustrative, say so.
• English timing in the jokes, Dutch directness in the claims. A joke arrives sideways. A claim arrives head-on. Get this backwards and it sounds like someone else.
• Contempt for the machine, never the operator. Punch at the dashboard, the platform, the rigged scoreboard. Never at the marketer who believed it.
• Concede, then reverse. Grant the premise fully, then turn the conclusion.
• Ask the question they cannot answer. Only when you genuinely do not know their answer. A rhetorical question you already answered is a sales technique wearing a toga.
• Story first, principle after. Open on the concrete thing, not the abstraction.
• Turn at the end. Let the last clause of a paragraph do the reversing.
• Land it and stop. A short final sentence. No summary after it.
• Ordinary words. A shop that sells dog collars, not "an affected SMB".
• Active voice. "We found", never "it was found".

Lines we use:

• "Meta says 4.2x ROAS. Google says 3.8x. Shopify says something else entirely. Which number do you tell your CFO?"
• "We don't do black boxes."
• "Once you see it, you can't unsee it."

Metaphors we own. Use these before inventing new ones:

• The formation. Channels have positions. Do not ask a defender to do the striker's job.
• The relevant chapter. One channel does not have to tell the whole story. It only has to be a relevant chapter.
• The traffic light. Each touch is green (it landed), orange (it helped, no love story) or red (wrong message or wrong audience).
• Marketing as an animal. A loyal pet that will eat you if it gets the chance.

Never use: best-in-class, world-class, industry-leading, cutting-edge, synergise, leverage (as a verb), optimise (as filler), potentially, might, could suggest, significant improvements, delve, in today's fast-paced, honestly, genuinely, to be fair, demand generation.

Audience-facing copy restores the cause-and-effect logic without naming the method. Category, research and investor copy may say "causal inference" by name.

How we decide

Three frameworks, used in one order. Higher wins. Every time.

1. What we are allowed to claim. Nothing below overrides it.
2. The One Thing (Keller and Papasan) decides what gets made. Write the One Thing of any piece in one sentence before starting. One job per page. One idea per piece. One amber per view.
3. Octalysis (Yu-kai Chou) decides why someone comes back. Name one primary drive per piece. Default to the top drives: meaning, accomplishment, creativity and feedback, ownership, social connection. Scarcity and loss only when the fact behind them is true and checkable. Curiosity with care: the answer has to follow.
4. Blitzscaling (Hoffman and Yeh) decides how soon it ships. We are at the family stage. Ship before you are comfortable. Let fires burn unless they touch the One Thing. Speed never buys a claim.

Conflict examples:

• Ship today, but the statistic cannot be traced: ship today without the statistic.
• Scarcity would lift sign-ups, but there is no real limit: no scarcity. Use a top drive instead.
• A second message is also true and also good: cut it. It becomes its own piece.

Design principles

1. Remove, then remove again. Cut until it breaks. Put back only what broke it.
2. One accent. Amber marks the one thing that matters in a view. Everything else is ink.
3. One type ramp. Six sizes. Nothing smaller than 11px.
4. Space before lines, lines before boxes.
5. Motion explains or answers. It plays out a cause, or it answers something the user did. Nothing animates just because it scrolled into view.
6. Say it once. If it is written somewhere already, point to it.

Visual system tokens

Colour

Retired: sky and violet. Do not use them.

Surfaces

Flat tones. No glass, no blur, no corner highlight. Never put one surface inside another.

Type

Geist for everything. Geist Mono only for code, hex values and units. Anton only for campaign and press work, uppercase.

Shape

• Pill: primary action only.
• 10px: cards and panels.
• 8px: inputs and tooltips.
• 2px with a flat 3px offset shadow in amber: press and campaign stamps only.

Motion

• Easing: cubic-bezier(0.22, 1, 0.36, 1).
• 120ms: press, toggle, icon swap.
• 200ms: hover, dropdown, tab, a chart value changing.
• 300ms: panel or dialog opening and closing.
• 8 to 20s, linear: ambient only, and rare.
• Everything stops for prefers-reduced-motion.

Charts

• ECharts with the causality theme.
• Series colours in order: #F5A623 (the series the chart exists to show), #A1A09E (nearest comparison), #676666 (everything else), #34D399 (went up), #FB7185 (went down).
• One amber series per chart. Grid at 4% ink, axis lines hidden. Tooltip on #1F1F22, 8px radius, value first.
• Series animate only when their data changes, 200ms. Never on load.
• Illustrative figures are labelled "Sample data" and never sit beside real ones.

Comedy

• North star: "Correlation isn't causation — and your marketing dashboard forgot."
• The law: a laugh is a safe surprise that is also true.
• The arc of every sketch: an absurd but real statistic, then the marketing mirror, then the reframe, then Causality Engine, signed off dry.
• The guardrail: the problem is the butt. Never the customer. The customer is the sane person in the room. The dashboard is the fool.
• Deadpan. Never explain the joke. No emoji laughter.
• Only real, verifiable data as the straight man.

Using BYQ Supply

We take structure and motion from BYQ Supply. We never take its look.

• Styles: measure, never import. Read the numbers in tokens.json. Never paste a BYQ DESIGN.md into our project.
• Wireframes: start here. Pick the structure before any styling.
• Components: keep the layout, replace every style with our tokens.
• Gems: at most one per page, and only when the motion explains a cause or answers the user. Keep every duration, delay and easing exactly as written, or do not use the gem. The essential colour becomes amber. Everything incidental becomes ink.
• Before shipping: no BYQ-hosted assets, only our tokens and type, reduced motion tested, licence checked, no href="#", and it actually runs in a browser.

Unresolved. Do not assert either way

1. Canonical background. This book uses #0A0A0C. Older product code uses #0F0F11. Use #0A0A0C, and do not call the question closed.
2. Em dashes. New copy in this book avoids them. Older chapters still contain them. Whether they are banned everywhere is open. Default to commas and full stops.
3. LinkedIn and Facebook format. Video or text first is not decided.
4. Customer fit floor. A minimum paid-media spend for fit has been proposed and not agreed. Do not state a minimum.
5. Hidden Value and Health. These appear in the book only as sample-data labels. Do not describe them as product features.
6. Analytics and indexing for this book. Not decided.

Changelog

• 7.3, September 2026. Rebuilt as the full book. Cinematic hero, numbered chapters, working tools inside the book (token playground, contrast checker, type tester, motion lab, voice linter, component gallery), How we sound, the positioning anatomy, the two ICPs, and this cheatsheet.
• 7.2, September 2026. Rebuilt around less is more. One type ramp, one accent. Added How we decide and Using BYQ Supply.
• 7.1, September 2026. What we say added. Positioning covers every channel, not just ads. Sample data labelled. Broken assets and analytics fixed.