Applying the Claude Code Hackathon Winner's Recipe?

[em3s]

Background

Actionbase is a database for serving user interactions (likes, follows, views, etc.) at scale. It currently handles million requests per minute at Kakao and was recently open-sourced. (Curious why we built this? See #32)

We received help from AI like ChatGPT and Claude during development.

AI Usage History

• 2023 Q2: GPT-4 — Project kickoff and initial design
• 2024 Q3: Claude 3.5 Sonnet — System stabilization
• 2025 Q1: GPT-4o, Claude 3.5 Sonnet — Building sustainable operations
• 2025 Q2: GPT-4o, Claude 3.7 Sonnet, Cursor — v3 reimplementation

Still, I was in control the whole time. Complex logic decisions, architecture choices, incident response—when hard problems came up, I was the one who had to solve them.

So for a while, I thought: "A single AI model can't fully replace me."

During the open-source preparation phase, I actually avoided using AI as much as possible. I wanted to see the overall situation with my own eyes.

What Changed

But very recently—just days after the open-source release—my thinking changed.

After going public, we started using it again. I got help with documentation and GitHub setup. @zipdoki built hands-on tutorials and a Go-based CLI to improve DX. What really shocked me was when @eazyhozy solved a core engine issue with Claude Code—a problem I thought only I could handle. It would have taken me a week to figure out alone. Opus 4.5-powered Claude Code was truly impressive. (We'll share a detailed write-up of this case separately.)

If this is how fast things are progressing, maybe it's better to ride this wave than stick to the old ways.

(That said, even with AI assistance, the final end-to-end verification still needs to be done by humans—that hasn't changed.)

The Recipe

As it happens, an Anthropic hackathon winner just shared their Claude Code setup:

👉 everything-claude-code

This repo contains agents, skills, hooks, commands, rules, MCP configs—all refined through 10+ months of actual use. It's not just a prompt collection. It's a workflow system.

What's Next

So we're going to try applying this to Actionbase development.

We don't know what the results will be yet. But we'll give it a shot and share what happens.

Questions:

• Is anyone already using Claude Code or similar AI agent workflows?
• What setups worked well, and what got in the way?
• If you've applied agent-based development to an open-source project, we'd love to hear about it.

[em3s]

Quick Update

Started experimenting with the everything-claude-code recipe!

• Ignore Claude Code config files for experimentation #92 (merged): Gitignore setup for experimentation
• Add Claude Code configuration (pilot/experiment) #91 (open): Claude Code config fine-tuned for our Kotlin/Java + Go stack

See #91 for usage instructions and available commands.