Google Engineers Reveal Go-To Prompts for Working with AI Models

Google engineers shared their go-to prompts for working with AI models, revealing a common pattern of using adversarial second opinions rather than agreeable assistants. Ten engineers and leaders were asked to share a prompt they personally refuse to work without, and the results showed that this approach leads to more robust planning and development.

The techniques used by these engineers include building specifications before writing code, making testing non-negotiable, running two-prompt cleanup passes, and grading one's own code like a harsh reviewer. These prompts were tested on a small task-tracker REST API, demonstrating how they can be applied in real-world scenarios.

One engineer, Maja Bilić, starts by assigning a skeptical persona to the model, forcing it to list key technical, UX, and architectural considerations before generating answers. This approach helps avoid over-engineering or oversimplifying solutions.