← ClaudeAtlas

context-graphlisted

Build and query a knowledge graph from someone's own markdown notes — extracting people, companies, projects and the relationships between them, so answers are grounded in their actual notes instead of guessed or retrieved by keyword. Use this when a user wants to make an Obsidian vault or notes folder queryable, asks questions spanning many notes ("who do I know at X", "which contacts relate to this project"), wants Claude to stop re-reading whole files to answer, or asks to build a knowledge graph, context layer, or semantic index over their notes. Runs a structural pass with a script, then a model pass that extracts entities and relationships, and answers later questions from the graph.
jessicahiggs/context-graph · ★ 1 · AI & Automation · score 67
Install: claude install-skill jessicahiggs/context-graph
# Context Graph Turns a folder of markdown into a queryable graph of **entities** (people, companies, projects, concepts) and **relationships** between them, then answers questions by retrieving the relevant slice instead of reading whole files. **Two passes, deliberately separated:** | Pass | Who does it | What it produces | |---|---|---| | **Structural** | `build_graph.py` | files, wikilinks, tags, headings, backlinks — deterministic, free, instant | | **Semantic** | you, the model | entities and relationships extracted from prose | The structural pass is cheap and re-runnable. The semantic pass costs model calls, so it is incremental: only notes whose content changed are re-extracted. ## Step 0 — Check before promising 1. **`python3`** — `command -v python3`. Ships with macOS and Linux; on Windows it may be `python`. 2. **A notes directory** the user can point at. No API key, no token, no scheduled job, no network access. Works on macOS, Linux and Windows. ## Step 1 — Ask what matters Two questions, no more: 1. **Where are the notes?** A path. Confirm it exists before running anything. 2. **What do they actually want to ask it?** This is the important one — it decides which entity types are worth extracting. Someone documenting an AI platform wants systems, models, datasets, metrics and owners. Someone doing research wants papers, authors and concepts. **Do not extract generic "entities"; extract the kinds they will ask about.** Agree a small type list, 4–6 at m