Writing Effective Prompts

Prompt Structure

Basic Formula

[Context] + [Specific Request] + [Output Format]

Example

• Context: “For the customer_orders table”
• Request: “show me data quality issues”
• Format: “in the last 24 hours”

Full Prompt: “For the customer_orders table, show me data quality issues in the last 24 hours.”

Components of a Good Prompt

1. Be Specific

   • ✓ “Show failed policies for customer_orders table”
   • ✗ “Show me some policies”

2. Provide Context

   • ✓ “As a data steward reviewing monthly compliance”
   • ✗ [No context provided]

3. State Your Goal

   • ✓ “I need to create a report for management”
   • ✗ [Unclear purpose]

4. Specify Output Format

   • ✓ “Provide as a table with columns for policy name, failure count, and asset”
   • ✗ [No format specified]

Context Provision

Why Context Matters

Context helps ADM:

• Select the right tools and agents
• Adjust tone and depth of responses
• Apply domain-specific knowledge
• Deliver accurate, relevant examples

Types of Context

Role Context

• “As a data engineer maintaining production pipelines…”
• “From a compliance officer’s perspective…”
• “As someone new to data quality…”

Temporal Context

• “For Q3 2025…”
• “In the last 24 hours…”
• “Since the last deployment…”

Scope Context

• “For all customer-related tables…”
• “Within the finance database…”
• “Across all data sources…”

Background Context

• “We recently migrated to cloud storage…”
• “Our policy requires 99% quality…”
• “This is for regulatory reporting…”

Specificity and Clarity

Vague vs. Specific

Use Concrete Terms

• ✓ “customer_orders table” ✗ “that table we use”
• ✓ “last 7 days” ✗ “recently”
• ✓ “policies with < 90% success rate” ✗ “bad policies”

Avoid Ambiguity

• Ambiguous: “Show me the data.” (Which data? Where? When?)
• Clear: “Show me the row count for the customer_orders table for each day in the last week.”

Examples

Question Answering

• Poor: “How’s the data?”
• Good: “What is the current data quality score for tables in the customer database?”
• Best: “Show me the data quality scores for all tables in the customer database, highlighting any below 95% over the last 30 days.”

Policy Creation

• Poor: “Make a policy.”
• Good: “Create a data quality policy for the orders table.”
• Best: “Create a comprehensive data quality policy for the orders table that validates:
  • Order amount is positive and less than $1M
  • Order date is not in the future
  • Customer ID exists in the customer table
  • Status is one of: pending, shipped, delivered, or cancelled”

Troubleshooting

• Poor: “Why did it fail?”
• Good: “Why did the customer_validation policy fail?”
• Best: “The customer_validation policy failed at 2 PM today with 450 rows. Analyze the failure pattern and suggest possible causes for the email validation rule failure.”

Analysis

• Poor: “Tell me about problems.”
• Good: “What data quality problems exist?”
• Best: “Analyze data quality trends for the finance database over the last 30 days. Identify:
  1. Most frequently failing policies
  2. Tables with declining quality scores
  3. New issues that appeared this month Provide recommendations for improvement.”