Advanced Prompting

Once you're comfortable with the basics, these techniques will help you configure agents faster and get more precise results.

Iterative refinement

Building with AI is a conversation, not a one-shot request. The best results come from an iterative approach:

1. Start broad — Describe the overall agent purpose or feature you want.
2. Review the result — Check what was configured.
3. Refine in follow-ups — Ask for specific adjustments.

Example sequence:

"Set up this agent to monitor my GitHub repos and alert on issues."

Review: The heartbeat is checking repos but only looking at PRs.

"Also check CI status on every heartbeat. Alert immediately on any build failures on the main branch."

Review: Good, but alerts are too noisy.

"Only alert on failures for the main and staging branches. Ignore feature branches. Batch PR updates into a daily digest."

Each step builds on the last, giving you fine-grained control.

Setting constraints

Tell the AI what not to do, as well as what to do. This prevents unintended side effects.

"Update the heartbeat to check every 10 minutes instead of 30. Don't change any of the existing skills or integrations."

"Add a revenue tracking skill. Don't modify the existing support desk behavior."

"Change the alert channel to #ops-alerts. Keep the daily digest going to #general."

Describing complex dashboards

For detailed canvas dashboards, describe the structure in sections:

"Build a dashboard with this layout:

• Top row: Four KPI cards showing MRR, Active Customers, Churn Rate, and Support Tickets.
• Middle section: A line chart showing weekly revenue trend for the last 3 months.
• Bottom section: Two tables side by side. Left table shows recent payments (customer, amount, date, status). Right table shows open support tickets (subject, priority, age)."

The AI handles multi-section dashboards well when you describe each section clearly.

Using roles and scenarios

Describe who receives what and in what context:

"Send P0 alerts to both the #incidents Slack channel and directly to the on-call engineer via Telegram. Send P1 alerts only to #incidents. Batch P2 and below into a daily digest to #engineering."

"Morning check-in should go to me on Telegram. Team standup summary should go to #standup on Slack."

Asking the AI to explain

You can ask the AI to describe what's configured or help you plan:

"What skills does my agent have right now? List them with what they do."

"I want to add incident response capabilities. Can you walk me through what that would involve before we start configuring?"

"What integrations is my agent connected to? Which ones are active?"

This is especially useful when you're picking up an agent after some time away.

Referencing existing behavior

"The heartbeat check for GitHub is working great. Set up a similar check for our GitLab repos using the same alert rules."

"Use the same escalation priority system from the support desk for incident alerts."

Multi-step features

For complex capabilities, lay out the plan first:

"I want to add meeting prep capabilities. Here's what I need:

1. Connect Google Calendar to pull my schedule
2. Research external attendees by looking up their company websites
3. Build a prep document canvas for each meeting with agenda and background
4. Track action items after meetings with owners and deadlines

Let's start with step 1."

Then work through each step in order, verifying as you go.

Usage-efficient prompting

Since billing is per-token, longer and more complex interactions cost more:

• Combine related small changes — "Change the heartbeat to every 10 minutes, add the staging branch to CI monitoring, and update quiet hours to 11pm-6am" is more efficient than three separate messages.
• Be clear the first time — Spending a moment thinking about your prompt saves usage on back-and-forth corrections.
• Use follow-ups wisely — "Also add..." is often more efficient than describing the whole configuration again.

Summary

Technique	When to use
Iterative refinement	Configuring any feature — start broad, then refine
Setting constraints	When you want changes scoped to specific areas
Describing complex dashboards	Canvas layouts with multiple sections
Using roles and scenarios	Agents with different alert targets
Asking for explanations	Understanding current config or planning next steps
Multi-step features	Large capabilities that need to be built incrementally