The Claude Code Setup Nobody Shows You (Replaces OpenClaw + Hermes)
The speaker explains how Claude Code can replicate and surpass tools like Hermes and OpenClaw using five core features: persistent memory, self-improving skills, interaction layers, scheduled tasks, and business context injection. The video argues that building your own setup inside Claude Code on a Pro/Max subscription is cheaper, more transparent, and more powerful than third-party agent frameworks. Business context is identified as the most critical and overlooked layer.
Summary
The video opens by critiquing the promise of agent frameworks like Hermes and OpenClaw — which pitch a simple loop of installing skills, scheduling jobs, and receiving results on your phone. The speaker acknowledges the appeal but argues that these tools are technically messy, expensive (requiring API credits beyond subscription), and lack deep business context awareness. Having gone through this frustration himself, the speaker built his own equivalent system entirely within Claude Code on a Pro/Max subscription.
The speaker identifies five core features that underpin any agentic system of this type. The first is persistent memory, which he breaks into four layers: (1) claude.md or agents.md as operating instructions, (2) a brand context folder holding voice, ICP, and positioning data, (3) an agent context folder with soul.md and user.md for personalization, and (4) project memory that tracks history and plans per project. He warns against bloating context files, advocating instead for short entry files that reference separate documents loaded only when needed.
The second feature is skills — reusable process documents that improve over time. He explains that Claude Code's built-in skill creator (made by Anthropic) can generate skills from descriptions or GitHub repos. He adds a self-learning loop by having skills request user feedback and store non-negotiable rules in a learnings.md file, so each skill improves with use.
The third feature is the interaction layer. The speaker pushes back on the common focus on mobile chat interfaces, arguing the real challenge is managing multiple concurrent goals and agents. He describes building a 'command center' — a Kanban-style UI wrapper over Claude Code terminals — that lets users supervise multiple business goals simultaneously, with sub-chats, plan views, and parallel agent tracking. For quick tasks, Claude Code's native Channels feature (supporting Telegram, iMessage, Discord) is sufficient.
The fourth feature is scheduled tasks. He explains that Claude Code can set up cron-style scheduling logic locally without needing a VPS. More importantly, he advocates chaining multiple skills within a single scheduled workflow — for example, pulling YouTube videos, analyzing them, and generating LinkedIn drafts every Monday morning. He emphasizes always including a human checkpoint before anything goes live, having moved away from fully autonomous publishing after finding 20% failure rates.
The fifth and most emphasized feature is business context. The speaker argues this is what Hermes and OpenClaw fundamentally lack out of the box — they don't know your brand voice, ICP, client details, or tone. He describes consolidating all business context into a single shared folder that every skill can reference, so updates propagate automatically. He calls this the 'compounding advantage' and the foundational layer without which all other features underperform. His closing advice is to start with the business brain before building agents or orchestration.
Key Insights
- The speaker argues that Hermes and OpenClaw, despite their polished demos, are just wrappers around the same five core features — and all five can be built natively inside Claude Code on a Pro/Max subscription, avoiding API credit costs entirely.
- The speaker warns that bloating context files causes 'context rot' — where output quality degrades the more context is loaded — and recommends keeping claude.md or agents.md concise, with separate reference files loaded only when needed.
- The speaker claims the real unsolved problem in agentic frameworks is not mobile chat access but managing multiple concurrent business goals in parallel, which existing frameworks like Hermes and OpenClaw fail to address by offering only one conversation at a time.
- After attempting fully autonomous scheduled workflows, the speaker found that roughly 20% of outputs were subpar, leading him to deliberately build in a human checkpoint before anything goes live — prioritizing quality supervision over full automation.
- The speaker identifies business context — not the agents or models themselves — as the true unlock for high-quality outputs, arguing that a shared brand context folder that all skills reference is the compounding advantage that third-party frameworks cannot replicate out of the box.
Topics
Full transcript available for MurmurCast members
Sign Up to Access