Gemini may finally leap out of the chatbox at Google I/O

While ChatGPT and Claude have been busy unleashing new models and eye-popping tools that unleash AI agents on code, your designs, and even your desktop, Gemini has been notably–perhaps even ominously, for its competitors–quiet.

Sure, there have been some interesting Gemini announcements recently, including its impressive Nano Banana 2 image generator, Lyria 3 (which cranks out music tracks based on your prompts), and a clutch of Google Workspace integrations. 

But as far as agentic coworking tools that live on your PC or autonomous AI agents that do your bidding, Gemini has seemingly lagged behind ChatGPT and, in particular, Claude.

I expect all that to change before the month is out.

The table is set for Google’s annual I/O conference, which is slated to kick off on May 19, and all signs point to a big surge for Gemini. Indeed, it feels like Google’s been holding its fire in terms of Gemini for the big I/O keynote, where we’re sure to see the tech giant peeling off one big Gemini announcement after another. Be sure to bring snacks.

One of Google’s possible Gemini announcements could be Proactive Assistant, a feature that (as I reported earlier this week) offers “personalized suggestions at the right time” based on data from your connected Google services (think Gmail, Google Calendar, and Google Drive), text messages, contacts, and even whatever’s on your screen. 

Even better, though, would be a rumored upgrade to last year’s Gemini Agent. According to Business Insider, Google staffers are already testing the update, code-named “Remy,” which it bills as a “24/7 personal agent for work, school, and daily life” that “elevates the Gemini app into a true personal assistant that can take actions on your behalf” as well as “handle complex tasks proactively.” 

On top of those developments, Google is widely expected to unveil–or at least tease–its next major Gemini model, likely Gemini 4. The new model would doubtless boast an agentic AI focus, and there are even whispers that it could natively support image and video generation. (Nano Banana 2 and Veo 3.1, Google’s current image and video models, are currently separate from the main Gemini model.)

In short, Google’s I/O conference could represent the culmination of the trend we’ve been seeing since the beginning of the year: AI assistants that leap out of the chatbox to help you, proactively, in the real world. 

And if Google does unleash its own 24/7 AI assistant, I’m sure there will be a pricey new subscription plan to go with it.

More in AI this week

• Anthropic has doubled five-hour Claude Code usage limits for Claude Pro and Max users, all thanks to a deal it struck with SpaceX for use of its Colossus 1 data center (as I reported earlier this week). That’s a bit surprising, given that SpaceX owner Elon Musk was calling Anthropic “misanthropic” and “evil” not all that long ago. 
• Turns out Chrome silently downloads a 4GB local AI model onto your desktop. Google says the tiny model enables a variety of helpful functionality, including the ability to summarize web pages and warn you about online scams, all without accessing the cloud. Don’t want the file on your drive? It’s easy to remove.
• I’m not a fan of emojis in AI chatbot responses, and ChatGPT has been–until now, anyway–one of the worst offenders. But with GPT-5.5 Instant, ChatGPT’s newest and fastest model, the flood of emojis should finally slow to a trickle.
• OpenAI’s new Codex AI pet is not only adorable, but it’s also surprisingly useful, notifying you when your AI coding agent is awaiting your approval or finished with a task. Every agentic AI app should have one of these.
• Specialized Claude AI agents now have a “dreaming” mode, a scheduled task that allows them to review and clean up the memories they’ve recently stored, as well as scour past chat sessions for “patterns and insights” (according to Claude’s API support site). 

Prompt of the week: The “Anti-Goal” Prompt

Why did you do that, ChatGPT? Didn’t I tell you not to rewrite my cover letter? Didn’t I ask ONLY for feedback on the letter I gave you? What gives?

You’ve probably had an exchange like that with an AI at some point, and the reason is simple: AI chatbots are so eager to make you happy, they often rush to complete your initial request while giving short shrift to any additional instructions (“…and do NOT rewrite my cover letter”).

Luckily, there’s a remedy: the “anti-goal” prompt. This structured prompt employs XML tags to ensure the AI sees and understands every part of your request, including the context (“here’s your role”), the goal (“here’s what to do”), and most importantly, the “anti-goal” (“here’s what NOT to do”). Give it a shot.

That’s all for now! 

Thanks for reading the latest issue of Prompt Mode. Want more next week? Don’t forget to sign up to start receiving this newsletter in your inbox soon. See you next time!