Weeknotes #64 (January 24, 2026)

Posted on: Jan 24, 2026

Things I found, read, did, and produced this past week.

Things I worked on

As a small project I was tinkering around, I was trying to see how well I could vibe code with assistance of Claude Code a CLI that would check if any css properties used in a stylesheet were not widely supported across web browsers: browser-compat-cli. An interesting tidbit on this was how Claude dug in, as I told it to explicitly use the caniuse-lite npm package, and it took 5-10 minutes while planning the implementation just exploring files in that package and writing sample node scripts to see example output. The context of that exploration was saved into the plan, but I had a thought how do you effectively save context for lesser known software libraries such that it doesn't need to do that exploration every time? I want to explore that in a future blogpost soon.
That said, I did a little digging and found on the frontpage of the caniuse website this problem was already fully solved: anandthakker/doiuse. So was my explorations with Claude for naught, yeah, but it was a low stakes CLI that I was making so it wasn't the end of the world that it was already built by someone else.

Digging around, I have been thinking now that coding agents can easily access CLI tools, what is the point of MCP anymore. Cause of this, I went looking and found this article from last August by Mario Zechner: MCP vs CLI: Benchmarking Tools for Coding Agents and found that in terms of a time and cost perspective for a benchmark he made, the difference is neglible. Also now that tools like Claude Code are not adding the full context of a MCP server on startup via their tool search tool similar to how it only loads skills when needed, I wonder if the software industry will keep using MCP or just return to using CLIs.
I finally tried out Andrew Nesbitt's git-pkgs, a tool that adds a git pkgs sub command to analyze your git history for dependency changes. I actually have a few use cases brewing in my mind already for this for some projects both personally and at work.
Maggie Appleton wrote a new article this week: Gas Town’s Agent Patterns, Design Bottlenecks, and Vibecoding at Scale. She comments how Gas Town definitely an out of control application and probably not the abstraction most people should use (which Steve Yegge also agrees with), but it is still useful to push these tools to their limits and down the road various more refined patterns will likely arise.