Brief: Laureo AI — personalized prompts + interactive capability tour (REVISED)

*Plain-language owner brief. Full build plan (for the agents) is in current-plan.md. Planning only — nothing built.*

What changed since you reviewed rev 1

You raised four things. I researched all four — three by digging through our own code, one by pulling in outside UX evidence — and rewrote the plan around them.

1. "I don't see the existing engine." You're right, and I was misleading. The smart prompt engine *is* built and wired — but it only appears inside a closed chat bubble you have to click open first (the "Ask Laureo" button, bottom-right of a company/deal/contact page). It's never shown on the main AI page. So you saw nothing because the bubble was closed — not because it's broken. The fix is the main AI page, which is what we're doing.

2. "Cache the instant prompts in the background, not at page load." Agreed and now the core of the design. We'll precompute each person's prompts ahead of time and store them in the database, so when the page opens it just *reads* them — no waiting, no AI call at load. There's already a feature that works exactly this way (the email "writing styles" + the dashboard's "AI Today"), so we're cloning a proven pattern: a nightly background job refreshes everyone's prompts cheaply, and a free rules-based version exists for every user with the smarter AI-written version reserved for active, paying users (keeps cost tiny).

3. "Add a personalized library section." Done — the "Browse all" library now opens with a "For you" section built from *your own org's real data* (real deals, contacts, overdue invoices), separate from the generic list.

4. The interactive onboarding tour. This is now a full pillar. A dismissible "See what Laureo can do" card (click to start, X to dismiss forever) launches a short 5-step interactive demo — "Step 1 of 5" — where the assistant actually shows itself drafting an email, creating a record, chasing an invoice, etc., on fake sample data (nothing touches your real CRM). It runs in both onboarding and on the AI page for first-timers.

Why this is the right call (the proof)

The research backs every decision:

• A Nielsen Norman Group study found 0% of users clicked AI prompt suggestions when they were generic and out of place — so ours must name your real records ("Chase the $8k Acme invoice, 12 days overdue"), sit right in the assistant, and say plainly what they'll do.
• Most people underuse AI because they can't envision what it can do ("the gulf of envisioning"; ~65% of US workers barely use AI) — which is exactly why the nudge-toward-capabilities tour matters.
• The closest products to us — HubSpot's Breeze (grounds answers in real CRM data) and Intercom's Fin (demos on your own data before you commit) — prove the thesis: real-data grounding is what makes it feel like more than a chatbot.
• Tours work best short (≤5 steps), skippable, and interactive — never a forced pop-up. That's how ours is designed.

The shape of the build

• One nightly background job + one small database table hold everyone's precomputed prompts (present before login).
• The 6 visible prompts are picked to span Laureo's "superpowers" (Know · Do · Send · Schedule · Get paid · Automate · Report) so you *see* it can act — and never show a prompt that would come back empty.
• A library drawer with "For you," job-category groups, and "Yours" (your pinned/custom prompts).
• The interactive tour, reusing our real draft/approval cards with scripted demo data — so it looks identical to the real thing but is free and safe.

Your three calls — locked

1. Sequencing → build everything together in one v1. The personalized prompts *and* the interactive tour ship as a single release.
2. Tour launch → opt-in dismissible card. A "See what Laureo can do" card; click to start, X to dismiss forever. No forced pop-up.
3. Tour demo data → scripted fake data only. Deterministic, free, and safe — nothing written to your CRM, no AI calls during the tour. Always works, even for a brand-new empty account.

The plan is final. I run /implement (isolated worktree, built end-to-end, then verify/test/ship) on your word.