Snowline Snowboards

This year, I made a whole new section of Snowline, my Japanese snowboarding site, to document all the Japanese Snowboard maker brands I could find as well. When I started this in January, I thought it would be like 15-30 brands with maybe 100 boards. It has spiraled a little out of control, totalling 70 brands and 770 models and 1890 editions logged.

Earlier this year, I did all the data collection by hand, and only used AI to build a tool to help me resize the images for the page. Only 8 months later, AI has gotten so good that I could point it to a URL of a board catalog (website, PDF, instagram posts) and it would figure out how to extract the line up for the year, compare it with last year's, make changes and extract the images.

When I first started the archive, I think I spent 3 weeks in total to build the site and input all the data. Now, I got the whole database updated to 26/27 season within 3 hours, with myself supervising and feeding it relevant URLs and instructions.

I needed to setup the right tooling for it, like creating an MCP server & CLI tools that contains image manipulation tools (both algorithmic and using ML models like Segment Anything Model (SAM)), methods to do operations on the database (query, update, add), instructions on how I mentally organized the database (brands, models, editions) and the target writing style (short phrases). Along the way, I made it regularly create skills for the project when it discovered new ways to edit the data and images or patterns it found in the data that I subconsciously had done when I first created it.

With that setup, I fire up Claude Code inside the directory and prompt it with things like this:

❯ Now let's update the ALC boards/editions. There are two sources 
  of information we should reconcile into our db, one is the PDF 
  catalog which is for 26/27 season https://alc-snow.jp/alc_26_27_collection.pdf
  and then there is their snowboard 
  website which shows the models and images. 
  https://alc-snow.jp/items.asp?cat=2485 
  Use both to figure out which models have changed or been retired or added, 
  also update the summary line for each model

  Read 4 files, ran 5 shell commands

Both sources agree on the line-up, with the main differences being that 
the original R-M appears on the website but is dropped from the 26/27 catalogue, 
and R-J switches from rocker to camber. Next, I'll compare the board graphics 
against what's in the library.

  Read 2 files, ran 2 shell commands

After each update, it will summarize what it did and what it needed input from me, I would double check the work. This is now doing a better job than if I did it, the images are higher quality and takes me so much less time to do it. This allows me to focus on designing a better site experience rather than data input.