Patterns

Recurring architectural and design patterns in AI/LLM systems — RAG, tool calling, routing, guardrails, and more.

Contents

• Retrieval-Augmented Generation (RAG) — Grounding LLM output with retrieved context
• Tool Calling & Model Context Protocol (MCP) — Universal standard for connecting LLMs to external tools and data
• Claude Tool Search Pattern
• Agent Skills Best Practices — Skill authoring, trigger design, permission model, validation checklist
• Fine-tuning Open Models — LoRA/QLoRA, Unsloth, axolotl, MLX, dataset prep, GGUF export for Ollama
• OpenClaw Workflow Prompt Library Pattern
• LLM Trust Boundaries Pattern
• Software Factories Pattern — Non-interactive development via seed, validation, and feedback loops
• Filesystem-as-Interface Pattern — Filesystem as the primary interface and persistence layer for agents
• Data Copilot MCP Tooling — Standardizing data access for Text-to-SQL pipelines using MCP
• Data Copilot Agentic RAG — Hybrid retrieval pattern for diagnostic analytics

Common Patterns

• RAG (Retrieval-Augmented Generation) — Grounding LLM output with retrieved context
• Fine-tuning — Adapting open models via LoRA/QLoRA for domain-specific behaviour
• Tool Calling & MCP — LLMs invoking external tools via structured schemas and the Model Context Protocol (MCP)
• Skills — Self-contained behaviour modules with triggers, instructions, and permissions
• Routing — Directing queries to specialised models or agents
• Guardrails — Input/output validation and safety filtering
• Chain-of-Thought — Structured reasoning prompts
• Multi-Agent Collaboration — Multiple agents coordinating on a task