5 Quick Links for Devs: Week 52, 2025
Replacing JS with just HTML
Aaron T. Grogg gives us a great refresher on why we should lean on the platform rather than reaching for a library. He walks through replacing common patterns—like accordions, autofilters, and modals—with native HTML elements like details, datalist, and the Popover API. It’s a solid reminder that standard HTML is often more performant and accessible than the JavaScript equivalents we've grown used to.
Karpathy Shocks the World
Andrej Karpathy set off a firestorm of reactions earlier in the week when he confessed that he feels like he's behind the curve when it comes to agent-based programming. What makes this confession so jarring is that just a couple of months earlier he was saying that agents weren't anywhere near ready for primetime, and that it would take another 5 - 10 years of development for them to become competent coders. Such a quick turnaround was spurred by the release of Anthropic's Opus 4.5 model, and the post itself set X ablaze with similar confessions. It feels like the average estimated timeline just moved forward half a decade..
Related: Claude Code creator says Claude wrote all his code for the last month
Scaling LLMs to larger codebases
Kieran Gill breaks down a practical framework for when you need to move beyond simple one-shot prompts. He outlines a "Research, Plan, Execute, Review" loop that focuses on managing context and creating a solid testing feedback loop. It's a much-needed guide on providing oversight and guidance to agents rather than just hoping for the best.
On LLMs in programming
Daniel Janus offers a "conscious LLM-skeptic" perspective that really resonated with me. While he admits these tools are useful and here to stay, he articulates the anxiety that comes with such a rapid shift in our fundamental workflows. It’s a thoughtful, slightly philosophical reflection on what we gain—and what we might be losing—as we hand more control over to the models.
Just Talk To It - the no-bs Way of Agentic Engineering
Peter Steinberger (steipete) cuts through the noise of complex prompt engineering with a simple philosophy: just talk to it. He argues that we spend too much time trying to "trick" the model with elaborated charades, when treating it like a capable partner—and giving it direct, high-level intent—often yields better results. A refreshing take on getting out of your own way.