Agent Best Practices

Start from the output you need

Tell the Agent whether you want a quick answer, a table artifact, a chart card, a brief, a dashboard, or model work.

Include the metric, time period, and grain

Good analytical prompts usually specify:

• Metric.
• Time period.
• Grain.
• Breakdown.
• Filter.
• Sort order.
• Output format.

Example:

Create a table artifact of net revenue and order count by channel by week for the last 12 weeks. Sort by the latest week's revenue descending.

Use governed names when you know them

Use the names you see in dashboards, Semantic Models, and the Catalog. If you are unsure, ask the Agent to discover available terms:

• "What revenue metrics are available?"
• "Which models include channel?"
• "What dimensions can I use with orders?"
• "Find the model that defines active customers."

Ask for lineage before trusting important numbers

For important metrics, ask:

• "Explain how this metric is calculated."
• "Trace this KPI upstream."
• "What is the grain?"
• "Which Transformation Model and Semantic Model does this use?"
• "Could this double-count because of joins?"

Prefer artifacts for work you will reuse

Use chat for quick exploration. Use artifacts when the result needs to persist:

• Table artifact for rows.
• Chart card for one visualization.
• Brief for narrative.
• Finding card for one insight.
• Dashboard for recurring reporting.

Break complex builds into stages

Instead of asking for a large finished dashboard in one message, use stages:

1. "Find the models and metrics that support this dashboard."
2. "Create chart cards for the key widgets."
3. "Build a private dashboard from those chart cards."
4. "Add filters and tabs."
5. "Explain lineage and dependencies."
6. "Prepare to publish."

Be explicit about filters

Ambiguous filters cause wrong answers. Say:

• Date range.
• Included or excluded statuses.
• Locations or regions.
• Channels or campaigns.
• Whether refunds, cancellations, tests, or corporate orders should be excluded.

Treat uploads carefully

When uploading CSV or XLSX files, always state:

• What the file represents.
• Whether it is one-time data or should match an existing source.
• Which columns are keys.
• Which columns are sensitive.
• Whether it can be used in a dashboard or only for ad hoc analysis.

Review before publishing

Before publishing dashboards, Transformation Models, or Semantic Models, ask the Agent to show:

• Dependencies.
• Preview results.
• Metric definitions.
• Downstream impact.
• Access assumptions.
• Any required confirmation phrase or publish review.

Good prompt patterns

Exploration

What data do I have access to for revenue analysis? Show the relevant sources, Transformation Models, Semantic Models, and dashboards.

Table artifact

Create a table artifact of top 25 products by net revenue for the last 30 days, with order count, margin, and refund rate. Mask any customer-level fields.

Chart card

Create a chart card showing weekly net revenue and order count for the last 90 days. Use the governed revenue metric.

Dashboard

Build a private dashboard for executive weekly performance. Include revenue, orders, AOV, active customers, channel performance, and a location table. Add date range and channel filters.

Transformation Model

Draft a Transformation Model that deduplicates customers by latest updated timestamp, joins order totals, and produces one row per customer. Preview it before publishing.

Semantic Model

Create a Semantic Model on the published customers_enriched transformation with active customers, new customers, country, region, signup date, and last seen date.

When to start fresh

Start a new conversation when:

• The Agent is carrying context you no longer want.
• You are switching to a different business question.
• You need to compare a clean answer against a previous analysis.
• You want to attach only specific artifacts to the next workflow.