When to use it

Use this prompt whenever a provider retires a model, you want a cheaper or stronger version, or a new release changes behavior. Model shutdowns are routine: Google’s documentation, for example, lists Gemini 2.0 Flash as shut down on June 1, 2026 with a recommendation to migrate to a newer Flash model. Teams that planned the move early had time to test; teams that did not found out when requests started failing.

It suits developers, product managers and technical leads who call a model through an API or build agents on top of one. It also helps non-engineers who need a structured list of questions to bring to their technical team.

How to customize it

  • [OLD MODEL] and [NEW MODEL] should include exact version names, since behavior differs between releases that share a family name.
  • [WHY I AM SWITCHING] changes the priorities. A forced shutdown favors speed; a cost-saving move favors careful quality comparison.
  • [FEATURES I DEPEND ON] is the most important field. Function calling, structured outputs, caching and long context are where silent differences hide.
  • [WHAT I ALREADY HAVE FOR TESTING] tells the assistant whether to start from scratch or extend an existing evaluation set. Our guide on building an evaluation set for LLM apps goes deeper.
  • [NUMBER] keeps the plan realistic for your team size.

Why it works

Most migration failures come from assuming the new model behaves like the old one. The prompt forces a list of differences before any code changes, and it asks the assistant to point to official pages instead of guessing, which avoids confident but wrong claims about specifications. The test-set step turns a vague feeling that the new model seems better into measurable numbers, and the rollout plan limits the damage if it is not.

Asking for cost per task rather than price per token matters too. A cheaper model that needs more tokens, or more retries, can cost more in practice.

Tips and variations

  • Paste your current system prompt and ask for the changes most likely to be needed for the new model.
  • Ask for a “migration risk score” per feature and sort the work by risk.
  • After the migration, rerun the plan every quarter as a lighter regression check; see LLM observability: what to log and monitor.
  • Never paste confidential customer data into a chat assistant unless your organization allows it; use anonymized examples.