A lot of the ideas in this piece on the new Ambiance harness seem obvious as I read the post, but not in a bad way. It’s more like giving voice to a few things that have been bouncing around in my head after having worked with LLMs, AI assistants, and, more recently, Claude code.
The central theme of the piece is that a very useful way to work with these things is to lean into what is well represeneted in the LLM’s training data which, in a Linux context, means allowing it to call handy and ubiquitous command line tools and, the author argues, to use the filesystem conventions that are already familiar to the model.
I have considered similar notions as I’ve evolved my own workflows with project-level planning docs and specs, home- directory-level notes for project-spanning endeavors, and skills that have been informed as to the quirks of where I store such things and how I like them to stay organized.
Where I’ve landed on the subject to date is that I definnitely agree that the LLM will be intimately familiar with the Unix-y filesystem layout, but I’m less certain that mere humans who haven’t been navigating the Linux FHS for 30 years are going to find /usr/share/docs a natural and intuitive place for notes and planning documents.
In any case, I like how the author is thinking and I’m looking forward to seeing how this pans out in practice.