Recently I’ve started working on the characterization data (chardata) app and the enclosing website using Claude to handle my workflow. It’s definitely addictive, but there are clear limitations and hazards.
I started work on the app as I needed something to do data extraction from CGATS files for another project I’m working on. In the distant past, I would have coded up something purpose-built for a single data extraction pattern in C++ (glued to my standard CGATS file reader class which I built eons ago), or for a one-off, extract the data in Excel after import as a demi-CSV file. I have to explain — by inclination I’m more of a back-end developer. Algorithms and functional code I’m pretty decent at, but front-end code is something which I never really liked doing. It shows on the admittedly lacklustre initial “minimalist blog” design of my initial blog site (https://colourbill.com). It’s more of a preference thing than anything else, tbh.
For this project, though, I was faced with dealing with many files each with their idiosyncrasies, and here’s me caught without a dev team for backup. What to do? AI to the rescue, of course.
It only took a couple of weeks to get to where I am now with the app (https://chardata.colourbill.com), but it took me less than a day to get to the initial working point where I could minimally load a CGATS file and show the results in a data table or 3d plot. This is the power of AI tools. (Claude in this case)
Within the first couple of days, I had not only the basic framework, but also a GitHub push workflow which enabled me to automatically deploy into the web (AWS instance). The manual bits (eg, figuring out how to do the handshake to get SSL set up within my domain) took longer than the time I actively spent commanding Claude Code to do its work.
Overall, the next couple of weeks was time I spent iterating and thinking on features, and doing things like figuring out how I wanted the mobile layout to work, when one of my beta testers commented that they couldn’t access the app because they were away from their laptop at the time.
By comparison, having worked with many other full teams doing traditional development in the past, as well as having slung code myself, I can say that the experience of doing development with a modern AI tool like Claude is 1 or 2 orders of magnitude faster. Sure, Claude is a bit like a junior developer — you have to be very precise in what you tell it, or else it’ll confidently give you something which it thinks is correct, looks on the surface to be correct, but is otherwise 100% wrong. In another post (Fogra 51 fails Fogra Media Wedge!) I talk about why this happens — basically if it’s not out on the Internet, Claude can’t figure it out from scant high-level description, and if it is on the Internet, you (the developer) should probably double-check to ensure that it grabbed the right info.
But acknowledging all that, I was able to create and iterate features which I was never able to get implemented before in my other color product because the ROI was too low. With AI, I can implement features so quickly that even though they are fairly small tweaks, I can easily justify them.
By the same token, this latter aspect is also perhaps the most insidious. It’s so easy to tweak, I managed to burn through a tonne of tokens before I knew it, and that’s with following best practices (compact regularly, separate chats to clear context, etc). But even then, using AI has been letting me go faster, bugs and fixes and all, than I ever could have before.
As Jensen Huang said, if your $500k engineer isn’t burning at least $250k in tokens, something is wrong. (https://www.instagram.com/reel/DWqy9RXEY-t/)
I’ll talk in another post about how I set up my current workflow, but I’ve been burning tokens simultaneously adjusting the website (still have a ways to go on this) as well as iterating the application. All quickly enough that I can afford to do what I wouldn’t have done in the past — specifically, taking time to set up my tools and environment so that my main work can go smoother. AI tooling is exponential — it lets you build the tools to build the tools to build the tools. Without it, I was just trying to put together high tech with bearskins and stone knives.