Prompting Guide

How you ask questions significantly impacts the quality of your results. This guide will help you write prompts that get you exactly what you need based on how Querri’s agent actually works.

Understanding How the Agent Works

Before diving into specific prompting techniques, it helps to understand the agent’s workflow:

Goal Setting - The agent translates your request into a specific goal
Data Discovery - Searches your Library for relevant sources
Selective Loading - Loads 1-2 most relevant sources (not everything!)
Investigation - Examines the data structure and content
Plan Creation - Designs steps to achieve your goal
Execution - Runs the steps and presents results

Knowing this workflow helps you ask better questions.

The Three Types of Data Interactions

The agent uses three distinct approaches depending on what you need:

1. Finding NEW Data (not yet in project)

When to use: You need data that isn’t loaded in the current project yet.

How the agent handles it: Searches your Library and automatically loads relevant sources.

Example prompts:

“Analyze my sales data”
“Load the customer database”
“I need revenue and cost data”

2. Investigating LOADED Data (already in project)

When to use: Data is already in your project and you want to understand it better.

How the agent handles it: Runs queries to examine the data structure and contents.

Example prompts:

“What’s in the sales_data step?”
“Show me statistics for the revenue column”
“What are the top customers by spending?“

3. Creating ANALYSIS (final outputs)

When to use: You want visualizations, aggregations, or transformations that users will see.

How the agent handles it: Creates and executes a sequence of analysis steps.

Example prompts:

“Create a bar chart of sales by region”
“Aggregate revenue by month”
“Forecast next quarter’s sales”

The Golden Rule: Be Specific

The more specific your question, the better the AI can help you.

Vague vs. Specific

Vague: “Analyze this data” Specific: “Show me monthly revenue trends for 2024”

Vague: “Look at customers” Specific: “Calculate the average order value for customers in the Northeast region”

Vague: “Clean the data” Specific: “Remove duplicate rows and fill missing values in the price column with the median price”

Specific questions give the AI clear direction about what you want to accomplish.

Structure of a Good Prompt

Effective prompts often include:

Action: What you want to do (show, calculate, create, find, filter)
Subject: What data or metric you’re interested in
Context: Time periods, filters, or conditions
Format: How you want to see the results (table, chart type, etc.)

Examples

“Show me [action] total revenue by product category [subject] for Q3 2024 [context] as a bar chart [format]”

“Calculate [action] the correlation [subject] between marketing spend and sales [context]”

“Filter [action] to orders [subject] where customer_type is ‘enterprise’ and order_date is after January 1, 2024 [context]”

Not every prompt needs all four elements, but thinking this way helps you be more precise.

Excellent Prompt Examples

Here are examples of well-crafted prompts organized by use case:

Data Exploration

“Show me the first 100 rows of the sales data”
“What columns are available in this dataset?”
“Give me summary statistics for all numeric columns”
“How many unique customers are in the database?”
“What’s the date range covered by this data?”

Filtering and Selection

“Filter to transactions from January 2024”
“Show only rows where status is ‘completed’”
“Give me customers in California, Texas, or New York”
“Select records where revenue exceeds $10,000”
“Find orders placed on weekends”

Aggregation and Analysis

“Show me monthly revenue trends for 2024”
“Calculate total sales by product category”
“What’s the average order value grouped by customer segment?”
“Count the number of orders per day in March”
“Sum revenue by sales representative”

Visualization

“Create a line chart of daily active users over time”
“Show me a bar chart comparing revenue by region”
“Plot customer acquisition by month as an area chart”
“Make a scatter plot of price vs. units sold”
“Display a histogram of order values”

Data Cleaning

“Remove duplicate rows based on customer_id”
“Fill missing values in the revenue column with zeros”
“Standardize all date columns to YYYY-MM-DD format”
“Trim whitespace from product names”
“Replace null values in the category column with ‘Unknown‘“

Forecasting

“Forecast next quarter’s sales based on historical trends”
“Predict revenue for the next 6 months”
“Generate a forecast for daily active users through year-end”
“Estimate customer growth for Q1 2025”

Advanced Analysis

“Find correlations between marketing spend, website traffic, and revenue”
“Identify customers whose spending has decreased by more than 20% compared to last year”
“Calculate customer lifetime value by acquisition channel”
“Show me the top 10 products by revenue growth rate”
“Detect outliers in transaction amounts”

Multi-Step Analysis Requests

You can ask for complex, multi-step analyses in a single prompt:

“Load the sales data, filter to 2024, aggregate by month, and create a line chart”
“Clean this data by removing duplicates and filling missing values, then show me total revenue by region”
“Calculate monthly revenue, identify the top 3 months, and forecast the next quarter”

However, breaking complex requests into separate messages often works better:

You can verify each step
You can adjust direction if needed
It’s easier for the AI to handle

Use your judgment based on complexity.

Time References

Be clear about time periods:

Good Time References

“for 2024”
“in Q3”
“from January to March”
“last 6 months”
“year-over-year”
“this quarter vs. last quarter”

Vague Time References

“recently” (when exactly?)
“this year” (calendar year or last 12 months?)
“historical” (how far back?)

Column and Field Names

When referencing specific columns:

If You Know the Exact Name

“filter where order_status is ‘shipped’”
“sum the total_revenue column”

If You’re Not Sure

“filter to completed orders” (AI will find the right status column)
“sum revenue” (AI will identify the revenue column)

The AI is good at inferring column names, but exact names help avoid ambiguity.

Asking for Help

When you’re not sure what to ask:

“What’s interesting about this data?”
“What analyses would you recommend for this dataset?”
“Help me understand customer behavior in this data”
“What patterns do you see in the revenue data?”

The AI can guide you toward meaningful questions.

Refining Results

If the first result isn’t quite right, refine iteratively:

Initial: “Show me revenue trends” Refinement: “Break that down by product category” Further: “Just show the top 5 categories” Polish: “Make that a stacked bar chart”

This conversational refinement is one of Querri’s strengths.

Common Pitfalls to Avoid

Too Many Questions at Once

Avoid: “What’s the revenue by month, which regions are performing best, are there any trends in customer behavior, and can you forecast next quarter?” Better: Ask these as separate questions

Assuming Context Not Yet Established

Avoid: “Now break that down by region” (when you haven’t created “that” yet) Better: “Show me revenue by region”

Overly Technical Language When Unnecessary

Avoid: “Execute a SELECT statement aggregating sum of revenue_column GROUP BY month_field” Better: “Show me total revenue by month”

You can use SQL-like language if you prefer, but natural language usually works just as well.

Iteration is Encouraged

Don’t worry about getting the perfect prompt on the first try. Querri is designed for conversation:

Start with a reasonable question
Look at the results
Ask follow-up questions to refine
Repeat until you have what you need

This iterative approach often leads to better insights than trying to craft one perfect mega-prompt.

Prompt Templates

Here are templates you can adapt:

Trend Analysis

“Show me [metric] by [time period] for [date range]“

Comparison

“Compare [metric] between [group A] and [group B]“

Top/Bottom N

“What are the top [N] [items] by [metric]?”

Filter and Aggregate

“For [filtered subset], calculate [aggregation] grouped by [dimension]“

Cleaning

“Remove [issue type] and [fix another issue] in [column or dataset]“

Forecasting

“Forecast [metric] for [time period] based on [historical range]“

Advanced: Understanding Agent Behavior

Goal Tracking and Deliverables

When you ask for something, the agent sets a goal and tracks deliverables.

Example: “Create a sales dashboard with 3 charts”

Goal: Build sales dashboard
Deliverables: Chart 1, Chart 2, Chart 3

The agent tracks what’s completed and what’s remaining. If you ask for something new mid-conversation, it becomes a new deliverable.

Multi-Turn Conversations

The agent can work across multiple turns (up to 30). It maintains context about:

Your original request
Previous steps created
Data currently loaded
Recent conversation

This allows natural refinement:

Turn 1: "Show me sales data"
Turn 2: "Filter to 2024"
Turn 3: "Create a monthly trend chart"

Two-Step Processing Pattern

Sometimes the agent needs to prepare data before creating final outputs:

Turn 1: Aggregate data by month (preparation)
Turn 2: Use aggregated data to create chart (deliverable)

This is normal and ensures clean, well-structured results.

When the Agent Asks Questions

If your request is ambiguous, the agent will ask for clarification:

“Which date column should I use—order_date or ship_date?”
“Did you want daily, weekly, or monthly aggregation?”

Providing clear answers helps the agent proceed correctly.

Working with the Library

Naming Your Sources

Give your data sources clear, descriptive names in the Library:

✅ “Q4 2024 Sales Report”
✅ “Customer Database (Prod)”
❌ “data_final_v3”
❌ “export_20241201”

Clear names help the agent find the right data faster.

Adding Descriptions

Add descriptions to your sources:

What the data contains
Date ranges covered
Any important notes

The agent reads these descriptions when searching for relevant data.

Using Tags

Tag related sources:

sales, revenue, quarterly
customers, crm
marketing, campaigns

Tags improve search and organization.

Common Agent Patterns

Pattern: Explore → Filter → Visualize

1. "Load sales data" → Agent loads and previews
2. "Filter to California" → Agent creates filtered dataset
3. "Create a bar chart by product" → Agent generates visualization

Pattern: Investigate → Decide → Create

1. "What's in the customer data?" → Agent examines structure
2. "Show me customer types" → Agent investigates values
3. "Create a pie chart of customer distribution" → Agent builds chart

Pattern: Compare Multiple Sources

1. "Load sales and costs data" → Agent loads both sources
2. "Join them on product_id" → Agent creates joined dataset
3. "Calculate profit margin" → Agent adds calculated column
4. "Show top 10 products by margin" → Agent creates result

Grounding in Data

If you ask “What’s the total revenue?” and the agent hasn’t analyzed that yet:

❌ Won’t make up a number
✅ Will create a step to calculate it from your data

This ensures all insights are grounded in your actual data.

Practice Makes Perfect

The more you use Querri, the better you’ll get at asking questions that yield great results. Pay attention to which prompts work well and adapt your style accordingly.

Pro Tips:

Start simple and build complexity through conversation
Let the agent load data selectively (1-2 sources at a time)
Review intermediate results before moving to next steps
Ask the agent to explain what it’s doing if unclear
Trust the agent’s judgment on data selection and analysis approach

Next Steps

Try these techniques in the chat interface
Learn about loading data effectively
Explore analysis capabilities
Review the Agent Tools Reference to see what’s possible