Anthropic's official ralph plugin gets the core principle of Ralph wrong by preventing a single instance of claude from exiting, rather than continuously starting new sessions.
Here's an opinionated, but proper implementation of the Ralph Wiggum AI technique.
Ralph solves the AI context window problem in the most simplistic way possible; keep feeding new AI sessions the same prompt until the job is done.
I'm not sure how they built theirs, but here's how I built mine:
- Parse the newsletter RSS feed
- Summarize the most recent post with an LLM in the style of a fast-talking 1920s transatlantic radio broadcaster
- Generate an audio file by synthesizing a custom voice I generated by sampling the recap narrator of Avatar: The Legend of Korra
- Upload the file to a cloud object store and return the playback URL to the client
- Repeat steps 1 - 4 each week until the heat death of the universe
This will run autonomously with new weekly content for as long as the newsletter persists.
I've used Koa in my last few work projects but will be either switching back to Express or moving onto Fastify for the next one. Like you mentioned, Koa does indeed feel dated. It's community never really took off, either; many of it's most popular helper/companion libraries haven't been updated in years.
I still use Koa out of habit since it was the only framework for a while that had first-class promise support.
One thing nice about Koa is that it's simple, so it's timeless in that way—it's not a moving target nor does it try to do something that needs a lot of core maintainers.