Hello, I'm the captain.
As you may know, this is a board for sharing experiences. I'd like to share some of my recent experiments and insights.
I wrote a slightly different version on my personal blog today (https://ghost.redsub.net/1in-gaebal-claude-code-sayonggi/). I've adapted the content for this board, focusing on 'development'. ㅋㅋㅋ
I primarily use Claude Code through a VS Code plugin rather than the CLI. The 'UI' is easier to read, and I occasionally open a terminal to run necessary tasks using the CLI. Plugin updates happen almost twice a day on weekdays.
Every morning after 8:00 AM, I brainstorm or plan my daily tasks with Claude Code. Since Claude Code has a 5-hour token limit, starting after 8:00 AM allows me to finish by 1:00 PM and use new tokens from 1:00 PM until 6:00 PM for two more sessions.
Until Opus 4.5, I could handle almost all requests with Opus 4.5 Thinking (high effort) without needing parallel tasks or multiple sessions. - Why would I need that for a one-person development project? - Tokens were sufficient.
The most important thing is that it's a one-person project. Parallel processing and quick returns are less crucial than using the time while AI works to review other things, brainstorm ideas, or organize additional tasks. As I use Claude Code more, I start scrutinizing the plans it provides meticulously. I often have to revise the plan persistently. This is more important than your CLAUDE.md. In fact, I sometimes overlook the content of CLAUDE.md. This isn't a matter of session-to-session memory retention; it's simply ignoring principles and making assumptions.
However, with the release of Opus 4.6, I started running out of tokens. The limit was being reached even before the scheduled time.
So, I began searching for plugins. I tried various tuned plugins, starting with Oh My Claudecode and applying them to my project.
Whether the code quality improvement is due to Opus 4.6 or the plugin is unclear. However, I discovered that 'token consumption didn't show a significant difference'.
Following Anthropic's guidelines, I decided to utilize CLAUDE.md, hooks, Rules, Skills, Agents, MCP, and other elements to their fullest potential. I also organized these elements in a MECE (Mutually Exclusive, Collectively Exhaustive) manner for optimal project development. I performed a clean installation of Claude Code and created a general-purpose plugin for one-person developers (https://github.com/redsub-captain/redsub-claude-code). Of course, I archived this repository today while writing this post.
The more I use Claude Code, the more I realize that 'tuning it properly for each project is crucial'. I've concluded that general-purpose plugins are not very meaningful. Of course, I still use the built-in Claude Code plugins effectively. The recent addition of Superpower is fantastic!
Therefore, I refined the context size and model selection (OPUS, SONNET, HAIKU) for each function within individual projects.
As a result, even with heavy workloads exceeding the 5-hour limit, I now only use up to 60% of the maximum limit.
Thanks to this optimization, I have some leeway within my MAX plan's $100 budget. After Opus 4.6 was released, I briefly considered whether I would need to upgrade to the $200 plan.
Of course, things may change as models evolve.
That's all for now. I just wanted to share a few thoughts in my spare time.
