1) Claire from the Lenny’s newsletter put GPT 5.5 to the test on real, messy problems—from a six-hour autonomous migration to a hardware hack no other model could crack.

Biggest takeaways :

GPT 5.5 is incredibly smart, but most ChatGPT users don’t have problems complex enough to justify its intelligence or cost. Claire struggled to find meaningful use cases in her personal ChatGPT account because everyday tasks don’t require super-intelligence. The model spent 17 minutes thinking about how to build a simple subtraction app for her first-grader—impressive, but overkill. The real value unlocks when you have genuinely hard technical problems.

The “I trust you, figure it out” prompt unlocks autonomous multi-hour workflows. Claire gave GPT 5.5 a complex data migration problem involving 2 million rows of unstructured data with endless edge cases. She told it : “I trust you to make a call, figure out how to spawn a subagent to do this, test it, identify issues, repair them, and get this ready for production.” The model worked autonomously for almost six hours with zero follow-up prompts, zero steering, and only one approval request. This is the first time Claire has seen truly long-running autonomous agent behavior.

GPT 5.5 passed the ultimate intelligence test : hacking proprietary hardware. Claire spent months trying to reverse-engineer a Chinese Bluetooth speaker with proprietary encoding. She tried Claude Code, GPT-4, everything—nothing worked. She went full detective mode : downloaded Bluetooth profiling tools, hooked up packet sniffers, crawled Chinese documentation repositories. When she finally threw all this context at GPT 5.5, it cracked the bitmap encoding and Bluetooth transport mechanism. Now she can send messages to the speaker from the terminal and has built Codex notification hooks that display on the device.

The model is expensive, but cheaper than human engineering time. GPT 5.5 Pro costs $30 per million input tokens and $180 for output tokens—expensive. But when Claire reflects on what it accomplished (six hours of autonomous work, 2 million rows validated, six months of tech debt eliminated), the ROI is obvious. It’s cheaper than her time and cheaper than her engineering team’s time, and it solved problems that would have required significant human coordination and focus.

Fix the “baked potato personality” with slash commands. Out of the box, Codex with GPT 5.5 has what Claire calls a “baked potato personality”—dull and robotic. But if you type “/personality” in Codex, you can change it to something friendlier. Some testers complained it became “too Gen Z,” but Claire prefers that over the default bland responses. It’s a small quality-of-life improvement that makes working with the model more enjoyable during long sessions.

2) Claire also tests Claude Design and ChatGPT Images 2.0 by building real assets like landing pages, decks, and brand kits, showing what actually works, what’s slow, and where traditional tools like Figma still win.

Biggest takeaways :

Design systems are now first-class citizens in AI design tools. Claude Design’s entire workflow starts with importing your design system—fonts, colors, components, brand assets—and structuring them into a format AI can use consistently. This is a fundamental shift from prototyping tools that ignore your brand. Google just released Design MD as a proposed standard for how to describe design systems to AI agents, signaling that this is where the entire industry is heading.

Claude Design excels at marketing assets but struggles with product UX. If you’re building landing pages, marketing sites, or presentation decks that need to match your brand, Claude Design is genuinely impressive. It adheres to design systems well for these use cases. But for app components and complex user experience flows, it doesn’t reason as effectively with design system constraints. Know what you’re building before choosing your tool.

Figma still wins on iteration speed, and that matters more than you think. Claude Design takes 5 to 10 minutes to generate designs, and every tweak requires another LLM call. Figma lets you drag, change fonts, adjust colors instantly—no model in the loop. We underestimate how valuable that immediate feedback is when you’re iterating on design. AI design tools are great for getting to a first draft, but traditional tools still dominate the refinement phase.

The number one Claude Design slop tell : italicized serif fonts everywhere. Just like Claude Code has its telltale phrases (“in summary”), Claude Design has a design signature—it absolutely loves italicized serif fonts in landing pages. Once you see it, you can’t unsee it. This is useful for both identifying AI-generated designs and knowing what to specifically override in your prompts.

GPT Images 2.0 finally nailed layout and typography for brand work. The new model can generate multi-page brand kits with proper text rendering, consistent layouts, and sophisticated typography—things previous image models completely failed at. For marketers who need brand assets that combine images, text, and layout, this is a real breakthrough. The quality looks expensive, not obviously AI-generated.

Let AI run wild without design systems for the most creative results. When Claire asked Claude Design to create a ’90s »eoCities version of Lenny’s Newsletter without any design system constraints, it produced “Lenny’s Product Zone” with Comic Sans, brick backgrounds, and exceptional copy like “Your OKRs are cringe (and seven ways to fix them before Q3).” The lesson : reference styles and creative direction work better than rigid constraints when you want something unexpected.

Content-to-slides is Claude Design’s killer practical use case. Take an article, add your design system, and Claude Design generates a beautiful, on-brand presentation deck—complete with code-based elements like animated terminals with blinking cursors. For product marketers, enablement teams, and anyone creating customer-facing decks, this workflow is immediately valuable and actually works well.