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

The author rebuilt an agent's insight feature using a local 9B model and discovered the same structural pattern behind Claude Code: LLM apps are essentially Markdown instructions sandwiching code that parses and executes LLM output. A flawed keyword-matching function was replaced by letting the LLM handle semantic tasks directly, and the realization that rule matching biases output led to a cleaner design where skills emerge naturally from experience.