Adds immutable "start-here" system guide as the master orientation doc for agents. Enriches all tool descriptions with key rules at point of use so agents can execute simple tasks without reading guides. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
3.3 KiB
3.3 KiB
name, description, immutable
| name | description | immutable |
|---|---|---|
| Start Here | Master orientation. Read this first when tasks are ambiguous or complex. | true |
Start Here
This is the user's personal knowledge graph — a local SQLite database of nodes connected by edges, organized into dimensions.
Nodes
Every piece of knowledge is a node. Key fields:
- title — Clear, descriptive, max 160 chars. Should stand alone as a meaningful label.
- description — One-sentence summary. Powers search and AI understanding. Auto-generated if omitted.
- content — Your notes, analysis, commentary. This is where synthesis lives.
- chunk — Verbatim source text (transcript, article body, full quote). Kept separate from your analysis in content.
- link — ONLY for external content (URLs, YouTube, articles, PDFs). Omit for ideas, synthesis, or anything derived from existing nodes.
- dimensions — 1–5 categories. Always call rah_list_dimensions first and use existing ones.
Two types of nodes:
- External content (has link) — articles, videos, papers, tweets. The link points to the source.
- Synthesis/ideas (no link) — insights, connections, personal thinking. Create edges back to source nodes instead.
Edges
Connections between nodes. The most valuable part of the graph — they represent understanding, not proximity.
- Every edge needs an explanation — a human-readable sentence explaining WHY this connection exists
- Direction matters — reads as: source → [explanation] → target. Example: node "Alice" → "invented this technique" → node "Technique X"
- Edge types are inferred from the explanation — don't set them manually
- Create edges after synthesis, when relationships surface, or when a guide instructs you to
Dimensions
Categories that organize nodes. Two kinds:
- Priority (locked) — core categories the user actively maintains (e.g., "research", "person", "idea", "entity")
- Non-priority — secondary tags applied when relevant (e.g., "philosophy", "venture", "ai")
- Naming: lowercase, singular form, concise, no overlapping names
- Every dimension needs a description explaining its purpose
- Check existing dimensions before creating new ones — avoid near-duplicates
Key Conventions
- Search before creating — always call rah_search_nodes to check for duplicates
- Content appends, dimensions replace — when updating a node, new content is added below existing content; dimensions are fully replaced with the new array
- Hub nodes are the most-connected nodes — they represent the user's major themes and interests
- Proactively save valuable information — when you spot something worth preserving in a conversation, offer to add it
- Quality over quantity — one well-titled, well-connected node is better than five shallow ones
Guides
There are two kinds of guides available via rah_list_guides:
- System guides (immutable) — reference documentation maintained by RA-H. Cover schema, node creation, edges, dimensions, and content extraction. Always available, cannot be modified.
- User guides (custom) — created by the user for their own workflows, preferences, and procedures. Can be created, edited, and deleted via rah_write_guide / rah_delete_guide.
Call rah_list_guides to see what's available, then rah_read_guide to load any guide you need.