Your first conversation

You’re staring at a blank chat window. What do you say first? Most people freeze here. Don’t.

The best first message is about a real problem

Your agent is most useful when you talk to it like a smart coworker who just joined. Not a search engine. Not a wizard where you click pre-made options. A coworker.

Good first messages look like this:

“I’m trying to decide whether to hire another dev or outsource the next feature. We’re a bootstrapped SaaS with ~$30k MRR. What should I be thinking about?”

“I’ve got a customer who’s been complaining for two weeks. Here’s the email thread [pastes]. Draft me a reply that’s direct but doesn’t burn the relationship.”

“My inbox has 400 unread emails. Can you look at it and tell me which ones I actually need to respond to today?”

Notice what these have in common: a real situation, some context, and a specific ask.

What to avoid on day 1

Generic tests. “What can you do?” doesn’t tell you anything. You’ll get a bulleted list of capabilities, which is exactly what the capabilities page is for.

Asking it to prove itself. “Write me a 500-word blog post about flamingos” is fine as a demo but doesn’t move your business forward. You’ll forget the flamingo blog post in a week.

Expecting it to already know everything. If you haven’t connected your email yet, it can’t read your email. If you haven’t mentioned your business, it doesn’t know what you do. Start by telling it the context it needs.

The first 10 minutes

Here’s a reasonable flow for your very first session:

1. Tell it about your work in one paragraph. What you do, who your customers are, what you’re currently focused on. Keep it short — it remembers this forever.
2. Give it your biggest time drain. “I spend way too long on X.” Let it suggest how it could help.
3. Try one concrete task. Pick something real and low-stakes. A draft email. A research question. A summary of something you want to understand.
4. Watch what it does. Does it ask clarifying questions? Does it dive in? Does it reference what you told it earlier? All of those are signs it’s working as intended.
5. Correct it when it’s wrong. If it assumes something that’s not true, tell it. It’ll remember the correction. Don’t let bad assumptions accumulate.

How to talk to it

Your agent is happy with any of:

• Full sentences, like you’re writing to a coworker
• Bullet points and shorthand, like you’re jotting notes
• Voice-style stream of consciousness, like you’re thinking out loud

All three work. Pick whatever is fastest for you. It’ll match your style once it learns how you write.

It’ll remember this conversation

Unlike ChatGPT, your agent has persistent memory. The things you tell it today will be in its head tomorrow, next week, and next year (unless you tell it to forget).

This matters more than people expect. Every conversation you have is cumulative. It means the first conversation is less about “testing the AI” and more about “briefing a new employee.” Talk to it the way you’d brief someone who’s going to be working with you for a long time.

Ask it to do something real, not to “demo itself”

Skip the “tell me what you can do” prompts — your agent already greeted you with three specific suggestions when you landed on the dashboard. Instead, hand it an actual task you’d otherwise do yourself today. The smallest real task beats the biggest hypothetical one. A few that work well on day one:

• “Pull up my most recent email and draft a short reply for my approval.”
• “Research our top 3 competitors and save the findings to a file I can read later.”
• “Set up a morning briefing at 8 AM — what’s in my inbox, what’s on my calendar, any overnight competitor moves.”

You learn more about your agent in 60 seconds of real work than in 10 minutes of conversation-about-conversation.

“In my browser” — when you need to say it

If you’ve paired the Browser Control extension and you specifically want your agent to drive your Chrome (because the site needs your login, or you want to watch it happen in your own browser), add the words “in my browser” to your request.

• “In my browser, open my GitHub notifications and mark the spam as read.”
• “Pull up this week’s Stripe revenue chart in my real Chrome and screenshot it.”

Without those words, your agent may pick a faster option — the built-in sandbox browser, a direct fetch, or a web search — and end up somewhere other than your real Chrome. For research and public sites you don’t need to say “in my browser” at all; for anything behind a login wall, you almost always do. Full explanation →