A couple of years ago I attempted and abandoned rebuilding this blog with more modern tooling and styles.
I wanted to revisit this and see how much effort it would take to update the blog given the explosion in the ability of agents over the last few months.
In my last attempt I set out the following criteria for rebuilding the tooling for this blog
Must
- Posts must be able to be written in markdown.
- Posts must support "dollar math" for maths and equations.
- The blog must fit with the Dask Branding Guidelines.
- The blog must have modern HTML metadata and preview images to perform well in search.
- Posts must be able to set the canonical header to point to original content.
- Anyone must be able to submit a new post easily without manual human interaction.
- Reviews must be performed in an open way such as in PRs on GitHub.
- Any Dask steering council member must be able to publish content to the blog.
- Readers must be able to subscribe via RSS.
Should
- The blog should be built using a tool familiar to Python developers.
- Posts should allow comments (the current blog uses Disqus).
- Posts should support custom JavaScript, CSS and HTML for when authors want to go off-road.
Could
- Posts could use MyST to reduce the amount of custom HTML that is used.
- Posts could be written in Jupyter Notebooks.
The approach I took in 2023 was to try to port the blog to Sphinx + ablog, given the amount of expertise in the Dask community around Sphinx. However it felt like the tool was fighting me at every turn and was part of the reason why I abandoned it.
This time I decided to explore multiple tools in addition to updating the blog to meet the requirements above. My goal was to port the blog to multiple frameworks and compare and explore what the project would look like when built with those tools. Then once I had a functioning blog with each tool I could explore updating the design for each one in a way that fit with the tool. This wouldn't have been possible without coding agents.
Last time I spent a day on it and got part way through updating the blog. This time I spent a day on it and created three different versions of the blog, all with new or updated frameworks, new updated designs and I even had time for some benchmarking using Lighthouse. I actually cloned the repo three times locally and had three instance of Claude Code running in parallel, one on each repo. This allowed me to switch between them and push them independently.
I'll open draft PRs for each one and get Netlify to run a preview build for each one so folks can compare them. The design of each one is slightly different due to the non-determinism of AI models. I'm not especially interested in comparing the design in any detail because any one of them can be prompted more to fix up any design issues. I'm mainly interested in comparing the different project structures, build tooling and framework features/limitations.
The design differences are interesting though. Each tool has taken a slightly different approach and each preview has different features. For example one added light/dark mode, another used cards for the index page instead of a list. I'm quite enjoying comparing the three and picking out things about the designs I like.
Once I've chosen a tool that I'm happy with I'll discard the other two and nudge the chosen one forward incorporating everything I've learned from this experiment and incorporating any design elements from the other two that I particularly liked.
A couple of years ago I attempted and abandoned rebuilding this blog with more modern tooling and styles.
I wanted to revisit this and see how much effort it would take to update the blog given the explosion in the ability of agents over the last few months.
In my last attempt I set out the following criteria for rebuilding the tooling for this blog
Must
Should
Could
The approach I took in 2023 was to try to port the blog to Sphinx + ablog, given the amount of expertise in the Dask community around Sphinx. However it felt like the tool was fighting me at every turn and was part of the reason why I abandoned it.
This time I decided to explore multiple tools in addition to updating the blog to meet the requirements above. My goal was to port the blog to multiple frameworks and compare and explore what the project would look like when built with those tools. Then once I had a functioning blog with each tool I could explore updating the design for each one in a way that fit with the tool. This wouldn't have been possible without coding agents.
Last time I spent a day on it and got part way through updating the blog. This time I spent a day on it and created three different versions of the blog, all with new or updated frameworks, new updated designs and I even had time for some benchmarking using Lighthouse. I actually cloned the repo three times locally and had three instance of Claude Code running in parallel, one on each repo. This allowed me to switch between them and push them independently.
I'll open draft PRs for each one and get Netlify to run a preview build for each one so folks can compare them. The design of each one is slightly different due to the non-determinism of AI models. I'm not especially interested in comparing the design in any detail because any one of them can be prompted more to fix up any design issues. I'm mainly interested in comparing the different project structures, build tooling and framework features/limitations.
The design differences are interesting though. Each tool has taken a slightly different approach and each preview has different features. For example one added light/dark mode, another used cards for the index page instead of a list. I'm quite enjoying comparing the three and picking out things about the designs I like.
Once I've chosen a tool that I'm happy with I'll discard the other two and nudge the chosen one forward incorporating everything I've learned from this experiment and incorporating any design elements from the other two that I particularly liked.