Prompts deserve a proper workflow.
Import prompts as maintained artifacts with explicit roles, declared variables, review history, repeatable tests, and a rollback path.
Read the guideAI Prompt Import API
Your prompts deserve a proper workflow.
Move beyond scattered snippets. Explore a prompt registry that preserves intent, makes changes reviewable, and separates reusable definitions from live executions.
JSONMESSAGES
Import definitions, not private executions
Give each prompt a purpose, stable identifier, declared variables, and message structure. Keep actual customer conversations separate from reusable instructions. Use reviewed synthetic examples when examples belong in the production definition.
A prompt named Ticket summary is not a unique identity. Define an owning collection or namespace so separate teams can maintain related artifacts without collisions.
Preserve roles and version behavior
Store the logical messages separately from model-specific rendering. Declare input variables and reject unresolved references. Keep substitution rules simple; imported prompt text should not execute arbitrary expressions or grant tool access.
Version output expectations, examples, and renderer configuration alongside wording when they affect behavior. An execution should identify the complete reviewed definition it used.
Review before activating
Separate received, validated, reviewed, and active states. Test resolved cases, ambiguous cases, and cases with missing information against explicit acceptance criteria. Valid formatting is not evidence of a faithful result.
Keep a known reviewed version for rollback and test the promotion path. Moving a prompt to a different model requires evaluation rather than an assumption of identical behavior.
Turn the idea into a test
Use a small, representative source and write down the expected result for one accepted record, one correction, and one repeated attempt. Keep the original values and the destination expectations visible throughout the review.
Use the format guides for source decisions and the local quickstart to see a simple mapping in action.
From Import Notes
A little reading. A better workflow.
Better context starts before the prompt.
Prepare a knowledge collection with meaningful chunks, source lineage, permission boundaries, and evaluation grounded in real questions.
Read the guide
Map it. Validate it. Understand it.
Separate transformations from validation, make every conversion explainable, and design useful errors around an explicit destination contract.
Read the guide