When we launched Dream Event, the headline feature was concept generation — describe your event, get a complete creative vision in minutes. But the feature we're most proud of is what happens next: the AI Event Designer.

The Designer is where you refine a generated concept through conversation. Change the color palette. Explore two different approaches to the cocktail hour. Tell the AI your venue just fell through and watch the entire concept adapt. It's the part of the product where AI and human taste actually collaborate.

Here's how we built it, what worked, and what we learned along the way.

Why a Chatbot Wasn't Enough

Our first instinct was to build a standard chatbot. One long conversation thread where users could ask questions and request changes. Simple, familiar, proven.

It was also terrible.

The Problem with Linear Chat

Event concepts have six or more interconnected dimensions — theme, programming, food and beverage, timing, décor, venue. A single conversation thread turned into chaos. Users would jump between topics unpredictably. The AI would lose context about what had been decided versus what was still open. Conversations hit 30+ messages and became impossible to navigate.

Worse, the AI didn't know when to stop asking questions. A user would say "I want a garden party" and the AI would respond with eight follow-up questions about flower varieties, seating arrangements, and lighting preferences. By question four, users were frustrated. By question eight, they'd abandoned the session.

We needed structure — but not so much structure that it felt like filling out a form.

Topic Mode: Structured Conversations That Feel Natural

The solution was what we call topic mode. Instead of one long conversation, the Designer breaks concept refinement into focused, scoped conversations — one per topic.

Six Core Topics

Every event concept generates six core topics, each with its own conversation thread:

Theme & Story Arc — the narrative your event tells from arrival to departure
Programming & Activities — what guests experience and when
Food & Beverage Design — menu concepts matched to the theme
Run of Show & Timing — how the evening unfolds minute by minute
Vibe, Décor & Environment — visual design, lighting, atmosphere
Venue Adaptation — making the concept work in a specific space

Each topic has a priority score. Theme comes first because it influences everything downstream. Venue adaptation comes last because it requires the other decisions to be in place.

Dynamic Topics

Not every event needs the same topics. The system detects signals in the user's brief and adds relevant topics automatically. Mention "immersive" or "interactive activities" and an Immersive Experience Design topic appears. Planning a conference? A Sessions, Talks & Facilitation Flow topic gets added. Overnight retreat? Lodging & Guest Logistics shows up.

This keeps the interface clean — you only see what's relevant to your event.

Option A/B: Giving Users Something to React To

Here's something we learned early: most people don't know what they want until they see it. Asking "what kind of cocktail hour do you want?" gets a blank stare. Showing two distinct options and asking "which direction feels right?" gets a clear answer.

How It Works

When the Designer generates a proposal for any topic, it presents exactly two alternatives — Option A and Option B. These aren't minor variations. They're genuinely different creative directions.

For a cocktail hour, Option A might be a structured tasting experience with guided pairings. Option B might be a relaxed open-bar setup with roaming small bites. The user picks one, asks to see more options, or describes their own idea entirely.

The UI renders these as clear, tappable buttons. Select Option A, request more options, or just type what you actually want. No constraints — the options are a starting point, not a cage.

Content vs. Operations

We learned that the A/B approach works brilliantly for creative decisions but feels forced for operational ones. Nobody wants to choose between "Option A: send invitations 8 weeks out" and "Option B: send invitations 6 weeks out."

So the system classifies each topic as content (creative ideation), operational (practical logistics), or balanced (a mix). Content topics lean into A/B proposals. Operational topics use direct, practical questions. This distinction is invisible to the user — the conversation just feels natural in both cases.

Knowing When to Stop Asking

The single hardest problem we solved wasn't generating good content. It was knowing when to shut up.

Clarification Question Caps

We implemented hard caps on how many clarification questions the AI asks before it must generate a proposal:

Content topics (theme, activities, décor): 4 questions max
Operational topics (logistics, timing): 2 questions max
Balanced topics: 3 questions max

After hitting the cap, the AI generates a proposal with whatever information it has. An imperfect proposal that the user can react to is infinitely better than a fifth question they won't answer.

Detecting Disinterest

Sometimes users don't want to engage deeply with a topic. They say "you decide," "whatever works," or "skip this one." Our system detects these signals and switches to strong defaults mode — it immediately stops asking sub-questions, provides a summary of what it recommends, and asks one final yes-or-no question.

This was a breakthrough for user retention. Before we added disinterest detection, users who hit a topic they didn't care about would abandon the entire Designer flow. Now they breeze through it in one message.

The First-Turn Rule

Another hard-won lesson: never ask detailed questions on the first message of a topic. When a user opens "Food & Beverage Design," the AI's first response is a 1-2 sentence summary of the current plan, followed by: "Let me know if you have any high-level feedback, and then I'll walk through some questions to nail the details."

This gives users an on-ramp. They can say "looks good, let's refine" or "actually, I want to go in a completely different direction" before getting into specifics. It respects their time and their agency.

Regeneration: The Concept Stays in Sync

When you finalize a topic conversation, the entire event concept regenerates. This is where the magic happens — and where we spent the most engineering time.

What Happens When You Lock a Topic

The system loads the current concept (the active version)
It gathers the full context: all prior versions, ad-hoc preferences you've expressed, change notes
It regenerates the creative brief, incorporating your topic decisions
It generates new visual concepts (mood board, spatial renders) aligned with the updated brief
It creates a new concept version — the old version is preserved, never overwritten

Every version maintains a lineage. You can always trace how the concept evolved from the original generation through each topic conversation. Nothing is lost.

Why We Never Mutate

Early on, we tried updating concepts in place. User finalizes the theme topic, we patch the theme section of the brief. Fast, simple, and it produced terrible results.

The problem is that event concepts are deeply interconnected. Changing the theme from "modern minimalist" to "rustic garden" shouldn't just update the theme section — it should shift the food and beverage direction, the décor, the suggested venues, the invitation tone. A patch approach misses these cascading effects. Full regeneration catches them.

The tradeoff is speed — regeneration takes longer than a patch. We mitigate this with focused prompting (the AI knows exactly what changed) and by running regeneration in the background while the user moves to the next topic.

Ad-Hoc Mode: When Users Break the Rules

Topic mode works for structured refinement. But users don't always want structure. Sometimes they want to say "change the event name to 'Midnight Garden'" or "add a photo booth to the programming."

That's ad-hoc mode — a freeform channel that lives alongside the topic conversations. The system analyzes each ad-hoc request and decides how to handle it:

Direct edits (rename, add an item, change a date): applied immediately as a targeted patch
Creative changes (shift the theme, redesign the menu): saved as a preference and applied during the next regeneration
Clarification needed: the AI asks a follow-up before acting
Brainstorming: the AI discusses ideas without modifying the concept

This dual-mode approach — structured topics for deep work, freeform ad-hoc for quick edits — gives users the best of both worlds.

What We Learned Building It

1. Constraints Make AI More Useful

Unconstrained AI conversations feel impressive in demos and frustrating in practice. Every constraint we added — topic scoping, question caps, disinterest detection, forced proposals — made the Designer more useful, not less.

2. Proposals Beat Questions

People are better at reacting than creating from scratch. Showing two options and asking "which one?" consistently outperforms asking "what do you want?" This isn't a limitation of our users — it's how human decision-making works.

3. Speed of First Value Matters

The time between "user opens Designer" and "user sees something useful" determines whether they stay. Our first-turn rule and aggressive question caps exist because every second of latency or friction is a chance to lose someone.

4. Version Everything

We almost shipped with mutable concepts. The decision to version everything — creating a new snapshot on every regeneration — felt like over-engineering at the time. It saved us repeatedly. Users change their minds. AI makes mistakes. The ability to trace and revert is invaluable.

5. Detect What Users Won't Tell You

Users rarely say "I'm bored of this topic" or "stop asking me questions." They just leave. Building detection for disinterest, confusion, and fatigue — and responding appropriately — was as important as the core AI capabilities.

What's Next

We're working on collaborative Designer sessions where multiple people can refine a concept together — imagine a couple planning their wedding in real time with the AI, each bringing their own preferences. We're also exploring visual previews generated during the conversation, so you can see how your decisions look before you commit.

The AI Event Designer is the feature that makes Dream Event feel different from a checklist generator. It's where AI stops being a tool and starts being a collaborator. And we're just getting started.

Want to see the AI Event Designer in action? Try Dream Event free — generate a concept and refine it in conversation.

How Our AI Event Designer Actually Works