Over the last year, as I've spent more time working with LLMs to power product features, I've also been thinking about how LLMs impact me as a long-form author. Of the ideas I've had, I'm particularly interested in the idea of building, and selling, books as "datapacks" for LLMs.

This time experimenting with MCP server for generating and running systems models ( github.com/lethain/syst... ). I think it's pretty wild how well this works, especially the tool to inject domain-specific language usage documentation into the context window.

Another writing workflow tool I've found useful is using LLM API to describe images. Here's is an example of script that replaces empty image descriptions with reasonably good ones. To extent the images could be better, it's b/c my prompt needs some work
gist.github.com/lethain/9c5e...

Small weekend project, I finally moved staffeng.com from Gatsby to Hugo. Gatsby is just fairly hard to maintain with so many JS dependencies for something I don't update frequently.

Now actually possible to fix issues again 😅

This one worked exceptionally well to create an SEO-driving FAQ page for page for new book. Done in a Claude workspace with entire book loaded into project via some light preprocessing plus repomix of all Markdown files.

So... should I make all the systems models look like Figma diagrams or Excalidraw diagrams? And why?

Now working on visual consistency of images for strategy book. One shot prompted a very simple script to build a static HTML page showing every image and descriptions used in the book. This is just a subset of the ~50 images overall, but gives a sense of starting point / problem to be solved.

Still enjoying my very basic, vibe coded typo/grammar fixer for blog posts 😂

I did make a few minor tweaks to reduce spurious regex matches by increasing context, and also to allow passing model in via "--model" parameter, but otherwise working well. (And I should have made those changes via Claude rather than by hand, but, ya know, still figuring out a good workflow.)

I'm still working on getting better at writing one-pass scripts using Claude. This one supporting a common blog editing workflow to incorporate grammar, typos, and other fixes (gist is gist.github.com/lethain/ce2a... )

Continuing to experiment with how foundational models can help with writing book _without_ using them to actually write anything (which is, ya know, what I want to do as the author). This was feedback on 180 pages, which is pretty good feedback and took <30 seconds to get versus many days.

One of my pointless personal goals was hoping the new book reaches 100 ratings on Amazon by end of 2024. If I miss, it will be by the narrowest margin 😂

Praying this is transient oddity so that I am not debugging my blog's build and deployment this evening 😆

Using latest is indeed my fault, but I've also had to debug end of life'd Ubuntu and Hugo versions. Entropy is real.

The professional books I read this year (I also read significantly more non-professional books but that's not really my lane).

Update on my "decade goals" for 2020s.