An experiment built entirely with Claude Code.
This project began as an experiment to see how far I could push Claude Code to build a complete application with minimal human intervention. I chose to create a compendium for the video game Hades, letting Claude handle everything from data scraping to deployment, building the database, building the frontend, deploying the app, debugging any errors that arose, and writing this readme.
I deliberately chose TypeScript for this project, thinking it would give the LLM the best chance of success. This had the added benefit of keeping me from tinkering with the code manually, as I'm not particularly comfortable in this ecosystem.
Throughout development, I observed several interesting patterns. By far the most common issue was Claude's tendency to use npm directly instead of deno, despite clear instructions in CLAUDE.md. Similarly, on several occasions, it would add raw CSS instead of utilizing Tailwind classes as directed.
What impressed me most was watching Claude iterate on its own solutions. Without any prompting from me, it would often realize a better approach and refactor its work. This autonomous improvement cycle was fascinating to observe. I was particularly impressed by its ability to iterate on troubleshooting problems even when I provided only vague descriptions of what was wrong.
There were also some minor technical issues—sometimes Claude would hang when trying to run the server, though this seems like a bug that will likely be addressed in future updates.
On a couple of occasions I gave up on trying to make it do what I wanted and did it myself because it was a quick fix. I made sure to note that by starting the commit message with "Mnaually ...".
I found it practical to use version control to create checkpoints whenever Claude produced a satisfactory solution. While Aider handles this automatically, I often prefer to squash or reword commits manually. The additional flexibility of Jujutsu over git became particularly useful.
I often struggled not to intervene with the choices Claude made or to resist the urge to dive deep into understanding all the generated code. This tension reminded me of a Wikipedia entry on Vibe coding:
If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding in my book—that's using an LLM as a typing assistant.
This project helped me realize that leaning into the vibe coding approach and allowing the AI to take the lead is a skill that requires deliberate practice and patience—one that I'm still developing.
The downside to letting LLMs handle all the work is that I genuinely enjoy coding myself. However, at least while the novelty lasts, there's some enjoyment to be found in watching computers code on their own.
The entire project cost approximately $10 in Claude API credits from start to finish.
See CLAUDE.md for instructions on running the server and processing data from the Hades wiki.