The base skill includes semantic vocabulary and layout recipes for LLM agents, RAG pipelines, memory subsystems, multi-agent orchestration, tool-calling loops, mind maps, and compact timelines.
| Concept | Node type |
|---|---|
| LLM or foundation model | llm |
| agent or orchestrator | agent |
| vector or embedding store | vector_store |
| working or short-term memory | memory |
| persistent store | database |
| tool or function | tool |
| API gateway | gateway |
Use arch-dark as the default for product and engineering diagrams. Use the Academic Overlay for paper-facing agent architectures.
| Meaning | Edge type |
|---|---|
| main request or response | primary |
| secondary data flow | data |
| trigger or control signal | control |
| read from memory | memory_read |
| write to memory | memory_write |
| non-blocking work | async |
| iterative reasoning loop | feedback |
Memory read and write share a color family but differ by line style, so the distinction remains visible in grayscale.
Arrange input, agent core, memory, tools, and output in a clear primary direction. Use control for tool invocation and feedback for tool results or replanning.
Separate write and read paths. Connect the memory manager to stores with memory_write, and stores to retrieval/ranking with memory_read. Use memory for volatile tiers and vector_store or database for persistent tiers.
Use layout: star for a single-level radial map. For deeper hierarchies, use hierarchical or explicit positions. Timelines are composed from horizontal nodes, modules, milestones, and dependency edges; there is no dedicated timeline engine.
Ready-to-render YAML lives under skills/drawio/references/examples/, including:
rag-pipeline.yamlagentic-rag.yamlmem0-memory-layer.yamlmulti-agent-orchestration.yamltool-call-loop.yamlAdd a compact legend whenever two or more edge semantics appear.