Designing Conversation Starters for Declarative Agents

You never get a second chance at a first impression: and that includes AI agents. Conversation starters are your agent’s handshake: the difference between a user who immediately sees value and one who types “hello” and bounces. Configuring them takes two minutes: getting them right is what this post is about.

Why Conversation Starters Matter More Than You Think

When someone opens your declarative agent for the first time, conversation starters appear as suggestion chips front and center in the chat: the most important piece of UX real estate your agent has.

Discoverability: Users don’t read documentation. Starters show them what your agent can do without requiring them to guess.
First-run experience: A blank chat is intimidating. Starters transform that empty void into a guided experience.
Capability showcase: Each starter is a window into a different feature of your agent. Five starters can demonstrate five different capabilities.
Reduced abandonment: When users know what to ask, they actually ask it. When they don’t, they leave.

I’ve seen agents with perfect instructions and access to all the right data: but zero conversation starters. Users typed “hi,” got a generic greeting, and never came back. All that work wasted because of a missing JSON array.

Don’t be that agent.

The JSON Configuration

Conversation starters live in your declarativeAgent.json file: an array of objects, each with a required text property (the message sent when clicked) and an optional title (the short label shown on the chip):

{
  "conversation_starters": [
    {
      "title": "What's my first-week checklist?",
      "text": "What's my first-week checklist as a new employee?"
    },
    {
      "title": "Who handles benefits questions?",
      "text": "Who do I contact for benefits questions?"
    },
    {
      "title": "Set up my laptop and accounts",
      "text": "How do I set up my laptop and accounts?"
    }
  ]
}

Keep title short and scannable. The text property is always required: if you omit title, the text is shown as the chip label.

💡 Tip

Use the text property to send a more detailed prompt than what fits on the chip. The title can be a short label like “PTO policy” while the text sends “What’s the PTO policy at Zava Insurance and how do I request time off?”: giving the agent more context to work with.

Design Principles for Great Starters

After building and reviewing dozens of agents, here are the principles that consistently produce high-engagement starters.

Show the Range

Each starter should highlight a different capability or topic area. If all five starters are about the same thing, you’ve wasted four opportunities to show what your agent can do.

For our Zava Insurance HR Onboarding Buddy, we want starters that map to distinct areas of the agent’s scope:

{
  "conversation_starters": [
    {
      "title": "What's my first-week checklist?",
      "text": "What's my first-week checklist as a new employee at Zava Insurance?"
    },
    {
      "title": "Who handles benefits questions?",
      "text": "Who do I contact for benefits questions?"
    },
    {
      "title": "Set up my laptop and accounts",
      "text": "How do I set up my laptop and accounts?"
    },
    {
      "title": "What's the PTO policy?",
      "text": "What's the PTO policy at Zava Insurance?"
    },
    {
      "title": "I need help with badge access",
      "text": "I need help with my badge access"
    }
  ]
}

The coverage spans onboarding logistics, benefits, IT setup, company policies, and support requests: a new employee scanning these chips instantly understands the agent’s full scope.

Be Specific, Not Generic

The most common mistake: starters like “Ask me anything” or “How can I help you?” These tell the user nothing about what the agent actually does.

Compare these two approaches:

Generic starters:

“Ask me a question”
“How can I help?”
“Tell me what you need”

Specific starters:

“What’s my first-week checklist?”
“Who do I contact for benefits questions?”
“How do I set up my laptop and accounts?”

The specific versions sound like natural conversation, not a feature menu.

📝 Note

Write starters from the user’s perspective, not the agent’s. Instead of “I can help with benefits enrollment” (agent voice), write “Who handles benefits questions?” (user voice). The chip should feel like something the user would naturally type.

Lead With a Verb

The best starters start with an action word:

“What’s…”: Asking for information
“How do I…”: Asking for guidance
“I need help with…”: Requesting support
“Who do I contact…”: Finding the right person
“Show me…”: Requesting a summary or overview

Verb-first starters create momentum: they signal the agent will do something useful.

Keep Them Short

Chips have limited visual space. Aim for 5-8 words on the title: enough to be specific, short enough to be scannable. Use text for longer prompts while keeping the chip clean.

Putting It All Together for Zava Insurance

Here’s the complete declarativeAgent.json for our HR Onboarding Buddy, with conversation starters complementing the instructions and description:

{
  "$schema": "https://developer.microsoft.com/json-schemas/copilot/declarative-agent/v1.6/schema.json",
  "version": "v1.6",
  "name": "HR Onboarding Buddy",
  "description": "Your friendly guide to everything onboarding at Zava Insurance.",
  "instructions": "$[file('instructions.txt')]",
  "conversation_starters": [
    {
      "title": "What's my first-week checklist?",
      "text": "What's my first-week checklist as a new employee at Zava Insurance?"
    },
    {
      "title": "Who handles benefits questions?",
      "text": "Who do I contact for benefits questions?"
    },
    {
      "title": "Set up my laptop and accounts",
      "text": "How do I set up my laptop and accounts?"
    },
    {
      "title": "What's the PTO policy?",
      "text": "What's the PTO policy at Zava Insurance?"
    },
    {
      "title": "I need help with badge access",
      "text": "I need help with my badge access"
    }
  ]
}

A new hire opening this agent sees five clearly labeled entry points covering the exact questions on their mind. No guessing: just click and go.

Once deployed, your conversation starters appear as clickable chips at the top of the Copilot chat, giving users a quick way to begin interacting with your agent.

The Value You Just Gained

A few lines of JSON bought you:

Immediate clarity: Users know what your agent does before typing a single word.
Higher engagement: One click lowers the barrier to that first interaction.
Capability discovery: Each starter is a gateway to a different feature.
Reduced support load: Users who find answers through starters don’t file tickets.
Professional polish: An agent with starters feels finished. One without feels like a prototype.

Two minutes of configuration. Five chips. A fundamentally better user experience.