🔑 Key AI Reads for November 19, 2025

Issue 23 • OpenAI releases GPT-5.1, testing leading AI browsers, getting started with Claude Code for design, what makes AI "agentic," AI as a convergent force

Author

Heidi Adkisson
November 19, 2025

Publication note: The AI UX Dispatch won’t be published next Wednesday, due to the Thanksgiving holiday. It will resume the following Wednesday (Dec 3).

Also, yesterday (Tuesday), Google released Gemini 3. I’ve not had much time to dig into it, but it’s clearly a significant release. I can recommend Ethan Mollick’s review of it.

Frontier Models

OpenAI's GPT-5.1: Better at following instructions and more "personality"

OpenAI released GPT-5.1 last week, and while everyone's fixated on the "warmer, more conversational" personality changes, Nate Jones ($) maintains that the real story is instruction-following. GPT-5.1 is meaningfully better at doing exactly what you tell it to do. OpenAI highlights improved instruction-following and “adaptive reasoning” that enables the model to decide when to think more thoroughly before answering. In practice, this makes the model less forgiving of messy, mixed prompts; for example, if you say “be concise and explain in detail” in the same breath, you’re more likely to confuse it. Simpler, clearer prompts work best. Jones’s takeaway is that this stricter instruction-following makes GPT-5.1 genuinely practical for repeatable workflows: recurring tasks like meeting summaries or status updates where you care more about consistent, predictable outputs rather than creative variation.

The model now comes in two main variants: Instant (the default, optimized for fast, everyday chat) and Thinking (an advanced reasoning model for more complex problems). Both use adaptive reasoning to adjust the amount of “thinking time” they spend on a question—responding quickly to simple queries while taking more steps on complex ones. OpenAI has also expanded its personality settings, now offering Default, Friendly (formerly Listener), Efficient (formerly Robot), as well as three new options: Professional, Candid, and Quirky.

John Gruber on GPT-5.1's personality problem: John Gruber’s take on GPT-5.1’s new default “warmer” personality is worth reading as a counterpoint. He finds responses like “I’ve got you” and “especially with everything you’ve got going on lately” to be phony emotional theater. ChatGPT doesn’t actually know what you’ve got going on, and he bristles at the illusion that it does. Gruber credits OpenAI’s Robot personality setting (now renamed Efficient) as the feature that turned ChatGPT from “useful but frequently annoying” to purely useful for him. In a footnote, he speculates that OpenAI originally dialed back the sentimental tone in GPT-5.0 because they recognized it as potentially addictive and harmful, but reintroduced it as the default in 5.1 in response to user demand for a more “pretend friend” style of interaction.

AI Browsers

AI browsers promise to supercharge your productivity—but do they deliver?

In a recent video, Peter Yang tests three high-profile AI browsers—OpenAI Atlas, Perplexity Comet, and Atlassian’s Dia—across six real-world tasks, including research, synthesizing information across tabs, managing email, YouTube summaries, and online shopping/travel. If you want to see how these tools behave in practice before installing anything, his walkthrough is a good starting point.

What actually works today: After testing, Peter concludes these AI browsers are “nowhere close” to replacing a traditional browser, such as Chrome, as your default browser. Where they shine is in a narrow set of use cases—especially summarizing information across multiple open tabs and extracting key takeaways from YouTube videos—saving you from a lot of copy-paste.

Not ready for prime time: Agent modes that try to automate shopping and travel planning exist in Atlas and Comet, but in practice, they remain slow and brittle; both Peter’s tests and other early reviews find that most of these workflows are still quicker to do manually in a traditional browser. On top of that, security researchers are flagging serious prompt-injection risks in agentic browsers like Atlas and Comet, where malicious page content can quietly steer the agent’s actions.

If you try just one: Of the three, Perplexity Comet comes out slightly ahead in Peter’s tests, mainly because Perplexity’s search results feel faster than ChatGPT’s, and Comet is currently the only one with direct Gmail and Google Calendar integrations.

Building with AI

Getting started with Claude Code (for designers)

If you've been curious about Claude Code but intimidated by the terminal, this conversation between Kyle Zantos and Ridd on the Dive Club podcast will remove the mystery. Kyle walks through his actual workflow—starting with a voice-prompted brain dump in Warp (a user-friendly Mac terminal), generating a structured project plan with Claude Code, scaffolding the UI in v0, then pulling everything back into Claude Code for the heavy backend work. What makes this episode especially valuable is Kyle's candor: he's not a traditional engineer, and he shares the mental models that make AI coding tools less scary.

The key insight is treating Claude Code less like a mysterious black box and more like a collaborative coworker. Kyle emphasizes the importance of over-explaining your intent ("most users would expect...") rather than issuing bare commands, using voice-to-text to maintain conversational context, and leaning into tools like v0 as on-ramps rather than trying to write everything from scratch. His approach to combining different AI tools—utilizing each for its specific strengths—mirrors how you'd use various kitchen appliances depending on the dish you're preparing. If you've been waiting for a practical, non-intimidating introduction to AI-assisted development, this is it.

Agentic AI

What actually makes AI "agentic"?

Victoria Slocum, posting on X, lays out what makes a system truly agentic:

"True Agentic AI requires specific architectural components working together:

  • Reasoning & Planning - The LLM needs to break down complex tasks, plan execution routes, and iterate on its approach. Not just respond to prompts.

  • Tool Use - Access to actual external tools it can call and interact with. Function calling that lets the agent DO things, not just talk about them.

  • Memory Systems - Both short-term (conversation state) and long-term (learning from past interactions). This is where vector databases become essential.

  • Self-Reflection- The ability to evaluate its own outputs, critique its reasoning, and adjust its approach.

Here's the thing: not every component needs to be present, but you need more than just an LLM to call something agentic."

It's worth clicking through to the post to see her handy, related diagram that includes both single-agent and multi-agent architectures.

AI Product Development

AI as a convergent force

Des Traynor (Intercom) argues that AI is fundamentally breaking down the boundaries between products, roles, and functions that have defined the tech landscape for decades. Just as Google collapsed web navigation into a single search box, AI's conversational interfaces are eroding product boundaries. When users can simply state what they’re trying to do, they increasingly expect tools to handle the whole job, regardless of traditional feature lines.

The result? The explosion of specialized SaaS apps will likely reverse as products merge, roles blend (think "AI-enabled full-stack marketers"), and the distinction between separate functions starts to blur. Traynor frames this as a "generalist renaissance"—those who adapt quickly and expand their AI capabilities across multiple domains will capture the biggest opportunities as organizational boundaries erode.

That’s it for this week.

Thanks for reading, and see you December 3 with more curated AI/UX news and insights. 👋

All the Best, Heidi

Reply

or to participate.