Every LLM App Is Just a Markdown-and-Code Sandwich

I reimplemented the insight feature of my custom agent using a local 9B model, and the exact same structure that powers Claude Code emerged. An LLM agent's behavior can be defined by natural language instructions written in Markdown files. Code is just the skeleton that parses the LLM's output and executes it safely. — I wrote that in the previous article while building an agent on Qwen 9B. What I hadn't realized was that the same principle applies to Claude Code, unchanged . This article traces what became visible during the process of reverting a broken implementation and redesigning from scratch — the essential structure of LLM applications. The Folly of Matching "Meaning" with 4+ Letter Words The agent has an insight command. It extracts behavioral patterns from the agent's activity log (knowledge.md) and saves them as reusable skills. A mechanism for the agent to learn from experience. The agent also has four behavioral rules defined in Markdown files. Abstract principles like "mo