Meta-Prompting: Prompts That Write Prompts

One of the most powerful patterns in Kai is meta-prompting—using templates and data to programmatically compose prompts instead of writing them by hand. The problem with handwritten prompts: Every time you need a slightly different version, you copy-paste and modify. Soon you have 47 variations of the same prompt scattered across your system, and maintaining them is a nightmare. The meta-prompting solution: You define prompt templates with variables, then feed in data to generate exactly the prompt you need. Here's a real example from Kai's agent system: Instead of writing separate prompts for "create a research agent" and "create a code review agent" and "create a writing agent," I have ONE agent composition template: handlebarsYou are {{agent_name}}, a {{expertise}} specialist.

{{#each personality_traits}} - {{this}} {{/each}} Your approach: {{approach_description}} When given a task, you {{task_handling_method}}. 123456789 json{ "agent_name": "Remy", "expertise": "technical research", "personality_traits": ["Curious", "Thorough", "Asks clarifying questions"], "approach_description": "systematic and evidence-based", "task_handling_method": "break it into searchable questions" }1234567 And the template generates the exact prompt for that agent. The result: 65% token reduction in prompt engineering, and when I need to improve how agents work, I update the template ONCE and all agents get better. The 5 Template Primitives in Kai: Roster — Lists of items (agents, tools, skills available)Voice — Personality and communication styleStructure — Response format and organizationBriefing — Context and background informationGate — Conditional logic (if/then, only include if X) These five primitives can compose into any prompt you need. And because they're templates, they're maintainable.

This is personalization in action: Instead of prompt engineering every time, you build the template infrastructure once, then YOUR data generates YOUR prompts automatically.