Teaching SHUR IQ to Write Like Us

The eyebrow (10px, uppercase, cobalt) names the category. The section title (28px Playfair) names the insight. The section intro (22px Playfair italic) gives the reader permission to keep reading by answering "why should I care?"

If the eyebrow and the title say the same thing, one of them is wrong.

If the title could be the name of a chapter in a book someone would actually read, it passes. If it sounds like a PowerPoint section divider, it fails.

The 240px ghost number (Playfair 900, 6% opacity) behind each section serves as a visual anchor. It should be a section counter ("01", "02") or a key metric ("96%", "$5.4B"). Never a word. Never decorative text.

Every tab in a multi-tab report follows the same structure: Hard Truth block (if warranted) → Hero section → Numbered sections. The reader should be able to predict where things are after reading one tab.

Never write "significant growth" when you can write "+14.3%". Never write "large market" when you can write "$14 billion." The number is the argument. The adjective is the crutch.

Use fragments to deliver the verdict. Use full sentences to lay out the evidence. The fragment hits first. The evidence follows for readers who want it.

Executives read conclusions. Analysts open the evidence. The <details><summary> toggle pattern puts the conclusion in view and the supporting data one click away. This keeps the page scannable while preserving depth.

This is the most common tell of AI-generated prose. Vary sentence openings. Start with numbers, names, verbs, or conditions. "The company" three times in a row is a machine pattern.

After writing, cut 30% of the default LLM output. Then read it again. If you can cut another 10% without losing meaning, do it. Density is respect for the reader's time.

Line 1: State what is broken. Be specific. Use numbers. Do not soften.

Line 2: State the opportunity hiding inside the problem.

Line 3 (pivot): Pose the tension the rest of the tab resolves.

If nobody flinches when they read it, it is not a hard truth. It is a summary. The Fiserv block made Josh Siegel uncomfortable. The Deckers block would make their CMO uncomfortable. That discomfort is the signal that the analysis went deep enough to matter.

"Here is what we found." "The question is whether..." "The risk is not that X. The risk is that Y." These work because they frame the report as investigation, not prescription. The reader's job is to evaluate what follows. The report's job is to present the evidence clearly enough to enable that evaluation.

In a multi-tab report, each tab's Hard Truth addresses a different facet. The Fiserv report: Tab 1 attacks brand identity, Tab 2 attacks product inertia, Tab 3 attacks organizational will. Each is uncomfortable for a different stakeholder. Together they create a comprehensive pressure map.

"[Company] has significant opportunities in digital transformation." Swap any Fortune 500 name in. Still true. Not a key message.

"Deckers has voluntarily dismantled every safety net it had." Only Deckers. Only this portfolio. Ownable.

When a call transcript exists, extract the exact words the client uses for their problems. Mirror those words in the report. If the CEO says "we have a storytelling problem," use "storytelling problem" — not "brand narrative deficit" or "messaging gap."

When no transcript exists, match the language in the company's public investor communications. They chose those words carefully. Use them.

Key messages compress three layers into a single readable line:

Stat: "96% of revenue from two brands."

Insight: "Deckers has no safety net."

Implication: "If either stumbles, there is nowhere to hide."

The best key messages combine all three: "96% revenue concentration means Deckers has no safety net — if either brand stumbles, there is nowhere to hide."

This is the single most common AI writing tell. "It's not a product — it's a platform." "It's not about the data — it's about the insight." Kill it on sight. Make the direct statement instead. "This is a platform." "The insight matters more than the data." Direct. No pivot. No rhetorical performance.

5 rules: Honest Framing, Narrative Structure, Anti-AI Voice, Condensed Hook, Audience Mirror. These are the base constraints. Every report gets them. Stored at: intelligence-brief/references/combined-v1-rules.md

CSS variables, typography scale, component specs. Ensures visual consistency. Stored at: intelligence-brief/references/design-system.md

2-3 canonical examples of the exact output format. The AI sees what "good" looks like before generating. This is where the annotated examples from this guide become training data.

When the team edits a report, the before/after pairs become correction data. Show the AI what it got wrong and what the fix looked like. 3 corrections are worth more than 30 rules.

The AI operates on pattern matching. Showing it "you wrote X, we changed it to Y" creates a direct mapping that abstract rules cannot achieve. The combined-v1 rules were themselves derived from a 5-branch experiment that generated hundreds of these correction pairs.

Focus meeting time on: "This line is wrong because..." and "This is what it should say instead." Capture those pairs. They compound.