Skip to main content

Copilot Best Practices and Limitations

This guide covers recommended practices for working with Coalesce Copilot effectively, along with its current limitations. Following these guidelines helps you get the best results while avoiding common pitfalls.

Copilot Is Evolving​

As with any AI-powered tool, Copilot is under active development. We regularly ship improvements based on feedback and testing, but you may occasionally encounter:

  • Unexpected Node configurations or SQL outputs.
  • Intermittent service interruptions.
  • Changes to available capabilities as we refine features.

Always Review and Verify​

Review Before Deployment

Examine all generated SQL, Node configurations, and transformations before deploying to production. Test changes in a development environment first.

Before deploying Copilot-generated work, verify:

  • Business logic matches your requirements
  • Generated code follows your organization's standards
  • Column mappings and transformations are accurate
  • Join conditions and relationships are correct
  • SQL syntax works in your target data platform
  • Node Types match intended use (such as Dimension, Fact, or View)
  • Business keys are properly set
  • System columns are preserved

Use Version Control Checkpoints for Safety​

  • Create a checkpoint before starting Copilot sessions
  • Review all changes in version control before deployment
  • Roll back easily if Copilot makes unwanted changes

Write Effective Prompts​

Be specific about requirements:

Instead ofTry
"Create a table""Create a dimension table for customers with customer_id as the business key"
"Add columns""Add columns for first_name, last_name, and email to the dim_customer Node"

Break complex tasks into smaller steps:

Start with basic structure, then add complexity iteratively. For example:

  1. First create the base dimension
  2. Then add calculated columns
  3. Then configure slowly changing dimension logic

Provide context and examples:

  • Reference existing Nodes: "Create a Node like dim_product but for stores"
  • Include domain-specific terminology
  • Specify Node Types when relevant (Dimension, Fact, Persistent Stage)
  • Include your data platform specifics when needed (Snowflake-specific functions, Databricks syntax)

Use iterative refinement:

  1. "Create a customer dimension"
  2. Review the generated Node
  3. "Add loyalty_tier and lifetime_value columns"
  4. "Set customer_id as the business key"

What Copilot Does Well​

  • Generating boilerplate transformations and staging layers
  • Converting SQL queries into Node chains
  • Adding columns and updating Node configurations
  • Explaining existing Node logic and dependencies
  • Building standard Coalesce patterns quickly
  • Rapid prototyping and iteration

Where You Should Be Cautious​

Exercise Extra Care

The following scenarios require additional review and testing:

  • Complex multi-node dependencies with intricate business logic
  • Migrations from legacy systems (always validate thoroughly)
  • Production-critical transformations without testing
  • Interpreting domain-specific business rules

Copilot Limitations​

Workspace-Level Operations​

Copilot cannot:

  • Create source Nodes
  • Create or delete entire Subgraphs (it can only add or remove Nodes from existing Subgraphs)
  • Modify Workspace settings or configurations
  • Manage Environments or deployments

External Data Operations​

Copilot cannot:

  • Query your data warehouse to preview actual data
  • Create new Storage Locations or database connections
  • Run actual data transformations (it only builds the DAG)

Advanced Customization​

Copilot cannot:

  • Set advanced Node configuration options
  • Create Custom Node Types
  • Modify Package definitions or install new Packages
  • Change Workspace-level templates or Macros

Data and Metadata Awareness​

Copilot:

  • Knows your table schemas and metadata
  • Doesn't know the actual data in your tables
  • Can't automatically detect data quality issues without your description
  • Can't recommend optimizations based on query performance

Team Collaboration​

  • Share Copilot-generated Nodes for peer review
  • Document business context that Copilot can't know
  • Use Copilot for scaffolding and the Coalesce UI for fine-tuning
  • Track time saved on repetitive tasks
  • Monitor accuracy of generated transformations
  • Identify patterns where Copilot excels or needs guidance

Security and Data Governance​

Understand Copilot's Scope​

Copilot Can AccessCopilot Cannot Access
Your Workspace metadata (Node names, columns, schemas)Actual data in your tables
Your Workspace permissionsExternal systems or databases directly
Your Storage Locations

Avoid Sharing Sensitive Information in Prompts​

  • Don't paste proprietary business logic that should remain private
  • Use generic examples when asking questions about patterns

Learn With Copilot​

Ask questions to deepen your understanding:

  • "How should I structure slowly changing dimensions?"
  • "What's the difference between Persistent Stage and View for this use case?"
  • "Can you explain the join logic you created?"
  • "What columns are in sales_fact?"
  • "Which Nodes depend on customer_dim?"

Troubleshooting​

If you encounter issues:

  • If Copilot stops responding or produces an error, wait 1-2 minutes before trying again
  • Start a new conversation if the issue persists. Sometimes a fresh context helps
  • Complex requests may take longer to process; simpler, focused prompts often work better
  • If results aren't what you expected, rephrase your prompt with more specificity or context
  • Report persistent issues to our support team so we can investigate

What's Next?​