For example, let’s say you’re planning an offsite and get an email from a team member asking for the hotel information so they can book a room. Ask Gemini in Gmail to look it up from a Google Doc that contains all the offsite details with: “what is the hotel name and sales manager email listed in @Company Offsite 2024.” Then, you can easily insert the information into your email reply.

4. Experiment and iterate to get better results

If at first you don’t succeed, try, try again! Refine your prompt and experiment with different approaches, including using synonyms or different keywords, adjusting the level of detail and specificity and testing different prompt lengths. (Based on the team’s research, the most successful prompts average around 21 words, yet people’s initial attempts are significantly shorter — usually fewer than nine words.) Plus, don’t forget you can ask follow-up questions.

Vishnu recommends experimenting with using different personas, too. For example, when you’re writing a prompt about training someone, you may want to ask Gemini to act as a colleague and then compare the results against asking Gemini to act as a teacher.

“When you’re happy with the results, take note of what’s working,” Vishnu says. “Try reusing some of that language with future prompts.”

5. Share what’s working — and what could work better

When development began on Gemini for Workspace tools, the team started a group chat where they shared examples of their prompts and Gemini’s responses with each other.

“The chat provided insights into the types of tasks people are using gen AI for, highlighting both its capabilities and limitations,” Vishnu says. “Googlers found success with practical tasks like summarizing itineraries and extracting information from documents.” As we’ve continued to upgrade our models, we’ve also been able to address previous areas of feedback, like how Gemini can contextualize information across multiple sources.

These are the kinds of insights the team is still gathering, both from users inside Google and from trusted testers. It all helps. As Vishnu puts it, “The more input we get, the more helpful we can make our products.”