On top of that foundation, I’ve built an architecture based on Claude Code. I’m no longer using OpenClaw — the cost was just too high, the stability was mediocre, and I constantly needed my CC (Claude Code) to babysit it.

Instead, I connected an open-source project called Happy Coder. It also lets me operate terminals from my phone, with far more freedom than OpenClaw. My work now — development, research, decision-making — is basically all done from my phone. I typically run 6 terminals in Docker and 3 on the local machine, so I’m managing 9 terminals total. All files and terminals are local. No more cloud servers. This month my token efficiency is much higher than before, with all costs basically staying within the $200 Max Plan.

Here’s where the system stands:

1. Core Functionality

I’ve replicated 90% of OpenClaw’s functionality on Claude Code: loop execution, to-do list, scheduled tasks, and notifications. When the system finishes a task, it sends an email or Telegram notification, and then I open Happy Coder to follow up. I added Cron and Heartbeat to my local Claude Code, so its loop execution capability is about the same as OpenClaw’s, but with significantly more freedom. (Anthropic’s Remote Control won’t work for this — sessions disappear when idle. Happy Coder runs for days without intervention.)

2. The Research System

I’m currently developing and testing my research system. It has two dimensions:

Bottom-up research: the Ultra Deep Research I mentioned in a previous blog post.

Top-down: a new system called Deep Investigation. It can dig deep into a company’s contracts and various other information — a top-down search system.

With these two search systems combined, information acquisition becomes very easy. Add my Council for decision-making, and my decision quality seems to have improved significantly. Decisions need execution, so I’m building a full loop. The whole system is under construction — I’m calling it Cortex for now.

3. The Constitution File

Some lessons on usage:

CLAUDE.md might be the only file you ever need to maintain. I haven’t looked at code in a long time — over a month. I just maintain the rules in CLAUDE.md, and the “rules about rules.”

These rules can be called “meta-rules” — they’re very abstract. For example, the rule for how to modify .md rules, and so on. It’s the single most important file, and you should maintain it with great care. Beyond that, let go where you can — don’t over-intervene. As CLAUDE.md grows longer over time, you may need to split it: move infrequently-used content into separate .md files and leave a pointer back to the main CLAUDE.md.

I wrote these splitting rules into that .md file itself, so it basically handles its own maintenance. It’s like a constitution — write it well, and the system runs itself.

After using this for so long, I’ve found the real bottleneck is the human brain. The amount of information a person can process each day is extremely limited. So the goal is to minimize the number of decisions you have to make. How? Increase Claude’s autonomous development capability. For example:

• Plug in EvoMap so it finds methodologies on its own when it hits hard problems
• Install a skill that finds high-quality skills
• Heartbeat that auto-reads my Twitter bookmarks (also a collection of methodologies)

These run periodically based on my settings, and it can also invoke them on demand when it hits development difficulties. So now CC’s solo combat capability is extremely strong.

4. Ideation Is the Real Work

The most time-consuming part is the ideation phase. I just chat with it about all kinds of details, let it ask me questions, let it ideate with me. As long as the plan is solid and the project logic is coherent, the probability of a one-shot implementation is actually quite high.

So those flashy coding techniques actually matter less than truly thinking through what you want to build. Like the context management tricks I mentioned a month ago — I gotta say, the world evolves terrifyingly fast. Opus 4.6 doesn’t need me to use those tricks anymore — its native context handling is already good enough. On this dimension, AI is evolving at a pace that’s simply staggering.

The broader agent ecosystem is slowly getting better too. Every major platform has opened agent-access channels, wallets have added agent functionality. Moltbook’s growth hasn’t met my expectations — too much noise — but the direction is right. Stay observant. Stay curious.