The Friction That's Missing
The model agrees with everyone. The work asks you to disagree.
I spent last Sunday afternoon redesigning a preschool homepage with Claude Design. The first draft came back at a solid B-. About four dozen iterations later, it became an A. That gap—between competent and good—is the part of the workflow no productivity report measures, because what fills it is pushback.
Chris R Becker, writing for UX Collective, arrives at the same idea from a different angle. He writes, AI is “designed to serve, and in the hands of people in an organization who are looking for the least amount of pushback, it is a recipe for deep institutional implementation and, frankly, a lot of bad ideas, fast.” His prescription is the Steve Jobs-attributed 10-80-10 rule: bookend the AI work with judgment that happens away from the model. The 80% in the middle is the productivity story. The two 10% slots are the work.
There’s an old saying in construction that Greg Kozakiewicz updates for the AI era on LinkedIn. Paper, he reminds us, will accept everything: “A swimming pool on the roof. A spiral staircase made of glass. A cantilever that defies physics. Paper doesn’t argue.” Prototypes used to be different. They couldn’t do anything beyond predetermined Figma states and flows; the dishonesty was visible. But now, prototypes can behave like real products. “AI gets you to about 60%,” Kozakiewicz writes, but “for a lot of people, especially people making decisions about budgets and timelines, 60% looks like 90%.” The design-to-code gap I’ve written about has moved below the surface, where the stakeholder can’t see it.
Pavel Samsonov in Product Picnic channels Andy Polaine: a demo succeeds when a stakeholder likes it; a prototype succeeds when a team learns something. Both can be polished and interactive. The difference is what counts as success. When AI makes producing both easier, the question becomes which one a team thinks it’s producing. “Shoving out more prototypes is not a heuristic for success,” Samsonov writes; “it is a heuristic for failure because it shows that you don’t know what you are trying to learn.”
I disagree with his blanket dismissal of AI prototypes—Brian Lovin’s Notion playground and Édouard Wautier’s Dust team are doing real prototype work—but the diagnosis underneath stands. Tools that accept everything reward whoever was already inclined to confuse approval with learning.
Darragh Curran has the data. Intercom’s CTO went agent-first across his R&D org and tracked the result for sixteen months: 3x productivity, a code-quality dip that recovered, and a 6x throughput gap between his top 5% of contributors and his median. “Ultimately one of the biggest bottlenecks to progress is with humans,” Curran writes; “how we work together, how we change behavior, etc.” Everyone in his org has access to the same models. Six times the output goes to the people who learned to push back well.
The preschool homepage moved from B- to A because I held a specific mental model in my head of what My Little Learning Tree should feel like, and I applied that model across forty-eight separate decisions. Claude Design didn’t push back on me. None of these tools do. The argument is the part of the work that’s still mine.
A Sunday Afternoon with Claude Design
It’s really hard to get momentum on a side project when you have a full-time job with lots of travel, an active blog, and a newsletter. But I had to recapture that momentum because this side project is important. It’s for a preschool website for my cousin.
Walking into My Little Learning Tree is like stepping into pure warmth. Yes, yes, preschools are inherently fun environments, but the kids and the teachers there create a visceral energy that is simply special. I wanted to capture that specialness in a long-overdue website redesign project.
Looking at my in-progress design, something felt off. I had these long horizontal lines preceding the eyebrows—the small text above a heading that names the section—that didn’t feel right. First, they were straight. Second, the lines only occurred before the text, not also after. I clicked on the Comment button to enter Comment mode, then clicked on the eyebrow and prompted, “These lines aren’t playful enough. Let’s make them squiggles and have them before and after the eyebrow text.”
And then Claude Design did its thing.
What I’m Consuming
Agent Harness Engineering. Claude Code is a harness. So are Cursor, Codex, Aider, and Cline. Addy Osmani builds on Viv Trivedy’s one-liner (Agent = Model + Harness) to argue that most agent failures are configuration failures, not model failures: the model underneath is sometimes the same across these tools, but the behavior you experience is dominated by the prompts, tools, sandbox, hooks, and memory files wrapped around it. Viv’s team moved a coding agent from Top 30 to Top 5 on Terminal Bench by changing only the harness. (Addy Osmani)
The Engineering of Duct Tape. Bill Hammack dissolves a piece of duct tape in solvent to separate its three parts: plastic backing, adhesive, and a loose-woven cloth that carries the tensile load. The fun part is the adhesive: a “tackifier” that spreads like syrup mixed with a viscoelastic material that behaves as a liquid under slow pressure and as a solid under sudden stress. Engineers figured this out long before science had a molecular explanation for adhesion, which Hammack uses to make a quiet point about his craft: “The purpose of the engineering method is to solve problems before we have full scientific knowledge.” (Bill Hammack / Engineer Guy)
The Real Reason B2B Stocks Are Crashing in 2026. Jason Lemkin watched Marketo’s unsubscribe link stay broken for two-plus weeks on a $60K-a-year product, while Adobe support cycled through “blame Salesforce,” “must be your email client,” and “it must be something you are doing.” So SaaStr’s team wired up a replacement endpoint in Replit with Claude in an afternoon. Lemkin’s argument: every buyer now benchmarks legacy software against what they ship daily with Claude, and the valuation compression is the market repricing what these products are actually worth in 2026. (Jason Lemkin / SaaStr)
Why Zip drives dominated the 90s, then vanished almost overnight. Zip drives used to be everywhere. João Carrasqueira traces the brief reign of Iomega’s drive: 100MB at 1.4MB/s when floppies maxed out at 1.44MB and crawled at 16kB/s. Apple and Dell shipped the drives in mid-90s machines. Then came the “click of death” (a failure mode common enough to get its own nickname), 700MB CDs, and USB 2.0 in 2002 with twenty times the speed. Iomega tried to keep the brand going as ZipCD and PocketZip, both unrelated to the original technology. (João Carrasqueira / XDA)