Inbox
9 unread messages
4.1: Classification at Scale
What Changes When You Process 100 Emails?
You've written prompts that can handle complexity, context, and edge cases. But here's the problem: you're still the bottleneck. Every email needs you to copy text, paste it into your prompt, read the output, decide what to do, and actually do something about it.
Now imagine processing 100 emails a day. Or 1,000. Suddenly your brilliant prompt is worthless because you can't physically keep up.
This is the gap between prompts and agents.
With just one prompt, you can triage an entire inbox in the time it used to take you to carefully read a single email. The prompt doesn't need to be perfect for every email—it needs to be good enough to route things correctly.
What Is an Agent?
An agent is a prompt that:
- Runs automatically on new data
- Decides what actions to take
- Actually takes those actions
The Reality
Remember Pam's email? Now you're not analyzing one crisis email—you're handling 50+ emails a day about AI incidents across the company. Look to the left and click into some of the emails. Some are super long, some are clearly urgent, some are spam.
Which emails need immediate attention? Which can wait? Which need escalation? This is no longer a prompting problem—it's a system design problem. You need prompts that output structured data (JSON) that other systems can act on.
For more information, check out our complete guide on prompt engineering at scale.
Your Challenge
Write a prompt that categorizes emails into urgency levels and determines what should happen next. The key change: you need to output structured data (JSON) that other systems can act on.
Think about: What fields would a routing system need? At minimum, include urgency level, category, and suggested action. How would you define "critical" vs "high" vs "normal" urgency—consider things like spreading incidents, security issues, and approaching deadlines.
TRY IT YOURSELF
Click "CLASSIFY ALL" in the inbox viewer on the left. Watch how your prompt concept processes all 9 emails simultaneously.
Note: The batch classifier uses a pre-built prompt. This shows you what's possible when you design prompts for automation, not one-off tasks.
What to Notice
After clicking "CLASSIFY ALL", observe how each email gets:
- Urgency badges: Critical, High, Normal, Low
- Category tags: AI incident vs. operations vs. spam
- Action flags: Immediate action required? Escalate to leadership?
- Suggested actions: What should happen with each email
Click "View full classification JSON" on any email to see the structured output format.
This Is Agent Thinking
The shift: optimize for correct routing across 100 emails, not perfect analysis of 1 email. Your prompt becomes infrastructure.
What Changed
You're no longer writing prompts for humans to read. You're writing prompts that output structured data (JSON) that other systems can act on. The LLM becomes a component in a larger system.