Skip to main content
System prompts are where you move institutional knowledge from your team’s heads into the agent. The data model tells the agent what your data looks like — system prompts tell it how to think about your business: what terms mean, which rules always apply, how to handle edge cases, and how to format outputs in a way that makes sense for your users. Agent Studio exposes three separate prompt fields, each targeting a different part of how the agent works.
Agent configuration panel showing the System Prompt, SQL Generation Prompt, and Chart Generation Prompt fields

​
Agent system prompt

The main system prompt shapes the agent’s overall behavior — its persona, its domain focus, and the business logic it should always apply. It’s the broadest of the three fields and the right place for: Business context and terminology Define the terms your team uses that don’t map directly to column names. If users ask about “revenue” but your underlying metric is gross_sales - refunds, say so here. If “active customers” has a specific definition in your business, encode it. Edge cases and business rules Describe situations the agent should handle in a particular way. For example: “When a user asks about quarterly performance, always use the fiscal quarter definition (Q1 starts February 1), not the calendar quarter.” Behavioral guardrails Set boundaries for what the agent should and shouldn’t do. For example: “Only answer questions about sales and revenue data. If a user asks about HR or payroll data, explain that this agent is scoped to revenue analytics.” Tone and response style Specify how the agent should communicate: level of detail in explanations, whether to proactively offer follow-up analysis, how to label charts, etc.
The system prompt is read on every query. Keep it focused — a tightly written 200–400 character prompt covering your most important rules will outperform a sprawling one that tries to cover every scenario. Use Golden Assets to handle specific question patterns instead.

​
SQL generation prompt

The SQL Generation Prompt is injected specifically when the agent is constructing a SQL query. Use this to encode rules about how queries should be written against your specific database — things that are too technical for the main system prompt but critical for query accuracy. Good candidates for this field:
  • Date/time handling rules (“Always use date_trunc('month', created_at) for monthly aggregations, not EXTRACT(MONTH FROM ...)”)
  • Schema-specific join logic (“The orders table must always be joined to customers via customer_uuid, not customer_id”)
  • Filtering defaults (“Always exclude test accounts where is_internal = true unless the user specifically asks about them”)
  • Performance hints (“Prefer filtering on created_at rather than updated_at for large table scans”)

​
Chart generation prompt

The Chart Generation Prompt is injected when the agent is building a visualization from query results. Use it to enforce formatting standards and chart conventions that should apply across all outputs. Examples of what belongs here:
  • Currency formatting (“If a field contains monetary values — price, revenue, cost, sales, profit — always format as USD with two decimal places”)
  • Axis and label rules (“Always include a date range label in the chart title when the query is time-bounded”)
  • List formatting (“When presenting bullet points or numbered lists in analysis text, do not add unnecessary line breaks between items”)
  • Color or chart type preferences (“Use bar charts for comparisons between categories; use line charts for trends over time”)
The chart and SQL prompts are additive — they layer on top of the main system prompt, not replace it. Put general behavioral rules in the main prompt and field-specific technical rules in the specialized prompts.

​
Next steps

Golden Assets

Add example queries and charts the agent references for common question patterns.

Data Model & Schema

Control which tables and columns the agent can access, and what they mean.