graphify

any folder of files → persistent knowledge graph → Obsidian vault, graph.json, audit report

/graphify ./raw

.graphify/
├── obsidian/        open as Obsidian vault — visual graph, wikilinks, filter by community
├── GRAPH_REPORT.md  what the graph found: god nodes, surprising connections, suggested questions
├── graph.json       persistent graph — query it weeks later without re-reading anything
├── cache/           per-file SHA256 cache — re-runs only process changed files
└── memory/          Q&A results filed back in — what you ask grows the graph on next --update

[placeholder: animated GIF showing the full pipeline — detect → extract → cluster → report → Obsidian vault]

Why this exists

The problem: Andrej Karpathy described it well: he keeps a /raw folder where he drops papers, tweets, screenshots, and notes. The problem is that folder becomes opaque. You forget what's in it. You can't see what connects. Ask Claude "what links paper A to the code in repo B?" and it will hallucinate — it hasn't read both, and even if it has, it has no memory of that connection next session.

What LLMs get wrong: Naive summarization fills in every gap confidently. You get a summary that sounds complete but you can't tell what was actually in the files vs invented by the model. And next session, it's all gone — no memory of what it extracted.

What graphify does differently:

• Persistent graph — relationships are stored in .graphify/graph.json and survive across sessions. Query weeks later without re-reading anything.
• Honest audit trail — every edge is tagged EXTRACTED (explicitly stated), INFERRED (call-graph or reasonable deduction), or AMBIGUOUS (flagged for review). You always know what was found vs invented.
• Cross-document surprise — Leiden community detection finds clusters, then surfaces cross-community connections: the things you would never think to ask about directly.
• Feedback loop — every query answer is saved to .graphify/memory/. On next --update, that Q&A becomes a node. The graph grows from what you ask, not just what you add.

The result: a navigable map of your corpus that is honest about what it knows and what it guessed.

Install

pip install graphify && graphify install

That's it. This copies the skill file into ~/.claude/skills/graphify/ and registers it in ~/.claude/CLAUDE.md automatically. The Python package and all dependencies install on first /graphify run — you never touch pip manually again.

Then open Claude Code in any directory and type:

/graphify .

Manual install (curl)

Step 1 — copy the skill file

mkdir -p ~/.claude/skills/graphify
curl -fsSL https://raw.githubusercontent.com/safishamsi/graphify/v1/skills/graphify/skill.md \
  > ~/.claude/skills/graphify/SKILL.md

Step 2 — register it in Claude Code

Add this to ~/.claude/CLAUDE.md (create the file if it doesn't exist):

- **graphify** (`~/.claude/skills/graphify/SKILL.md`) — any input to knowledge graph. Trigger: `/graphify`
When the user types `/graphify`, invoke the Skill tool with `skill: "graphify"` before doing anything else.

Usage

/graphify                          # run on current directory
/graphify ./raw                    # run on a specific folder
/graphify ./raw --mode deep        # more aggressive INFERRED edge extraction
/graphify ./raw --update           # re-extract only changed files, merge into existing graph
/graphify ./raw --watch            # notify when new files appear

/graphify add https://arxiv.org/abs/1706.03762        # fetch a paper, save, update graph
/graphify add https://x.com/karpathy/status/...       # fetch a tweet
/graphify add <url> --author "Karpathy" --contributor "safi"

/graphify query "what connects attention to the optimizer?"    # BFS — broad context
/graphify query "how does the encoder reach the loss?" --dfs   # DFS — trace a path
/graphify query "..." --budget 1500                            # cap at N tokens

/graphify path "DigestAuth" "Response"      # shortest path between two concepts
/graphify explain "SwinTransformer"         # plain-language node explanation

/graphify ./raw --html             # also export graph.html (browser, no Obsidian needed)
/graphify ./raw --svg              # also export graph.svg (embeds in Notion, GitHub)
/graphify ./raw --neo4j            # generate cypher.txt for Neo4j import
/graphify ./raw --mcp              # start MCP stdio server for agent access

Works with any mix of file types in the same folder:

Type	Extensions	How it's extracted
Code	.py .ts .tsx .js .go .rs .java .c .cpp .rb .cs .kt .scala .php	AST via tree-sitter (deterministic) + call-graph pass (INFERRED)
Documents	.md .txt .rst	Concepts + relationships via Claude
Papers	.pdf	Citation mining + concept extraction
Images	.png .jpg .webp .gif .svg	Claude vision — screenshots, charts, whiteboards, any language

What you get

After running, Claude outputs three things directly in chat:

God nodes — highest-degree concepts (what everything connects through)

Surprising connections — cross-community edges; relationships between concepts in different clusters that you didn't know to look for

Suggested questions — 4-5 questions the graph is uniquely positioned to answer, with the reason why (which bridge node makes it interesting, which community boundary it crosses)

The full GRAPH_REPORT.md adds community summaries with cohesion scores and a list of ambiguous edges for review.

Key files explained

File	Purpose
GRAPH_REPORT.md	The audit report. God nodes, surprising connections, community cohesion scores, ambiguous edge list, suggested questions.
graph.json	Persistent graph in node-link format. Load it with NetworkX or push to Neo4j. Survives sessions.
obsidian/	Wikilink vault. Open in Obsidian → enable graph view → see communities as clusters. Filter by tag, search across everything.
.graphify/cache/	SHA256-based per-file cache. A re-run on an unchanged corpus takes seconds.
.graphify/memory/	Q&A feedback loop. Every /graphify query answer is saved here. Next --update extracts it into the graph.

What this skill will NOT do

• Won't invent edges — AMBIGUOUS exists so uncertain relationships are flagged, not hidden. If the connection isn't clear, it's tagged, not fabricated.
• Won't claim the graph is useful when it isn't — a corpus over 2M words or 200 files gets a cost warning before proceeding.
• Won't re-extract unchanged files — SHA256 cache ensures warm re-runs skip everything that hasn't changed.
• Won't visualize graphs over 5,000 nodes — use --no-viz or query instead.
• Won't download datasets or set up infrastructure — graphify reads your files. What you put in the folder is what it works with.
• Won't implement baselines or run experiments — it reads and maps. Analysis is yours.

Design principles

1. Extraction quality is everything — clustering is downstream of it. A bad graph clusters into bad communities. The AST + call-graph pass exists because deterministic beats probabilistic for code.
2. Show the numbers — cohesion is 0.91, not "good". Token cost is always printed. You know what you spent.
3. The best output is what you didn't know — Surprising Connections is not optional. God nodes you probably already suspected. Cross-community edges are what you came for.
4. The graph earns its complexity — below a certain density, just use Claude directly. The graph adds value when you have more than you can hold in context across sessions.
5. What you ask grows the graph — query results are filed back in automatically. The corpus is not static.
6. Honest uncertainty — EXTRACTED, INFERRED, AMBIGUOUS are not cosmetic labels. They are the difference between trusting the graph and being misled by it.

Contributing

Adding worked examples

Worked examples are the most trust-building part of this project. To add one:

1. Pick a real corpus (people should be able to verify the output)
2. Run the skill: /graphify <path>
3. Save the full output to worked/{corpus_slug}/
4. Write a review.md that honestly evaluates:
   • What the graph got right
   • What edges it correctly flagged AMBIGUOUS
   • Any mistakes or missed connections
   • Any surprising connections that were genuinely surprising
5. Submit a PR with all of the above

Improving extraction

If you find a file type or language where extraction is poor, open an issue with a minimal reproduction case. The best bug reports include: the input file, the extraction output (.graphify/cache/ entry), and what was missed or invented.

Adding domain knowledge

If corpora in your domain consistently contain structures graphify doesn't extract well (e.g., legal documents, lab notebooks, musical scores), open a discussion with examples.

Worked examples

Corpus	Type	Reduction	Eval report
Karpathy repos + 5 research papers + 4 images	Mixed (code + papers + images)	71.5x	worked/karpathy-repos/review.md
httpx (Python HTTP client)	Codebase	—	worked/httpx/review.md + GRAPH_REPORT.md
Mixed corpus (code + paper + Arabic image)	Multi-type	—	worked/mixed-corpus/review.md

Each includes the full graph output and an honest evaluation of what the skill got right and wrong.

Tech stack

Layer	Library	Why
Graph	NetworkX	Pure Python, same internals as MS GraphRAG
Community detection	Leiden via graspologic	Better than K-means for sparse graphs
Code parsing	tree-sitter	Multi-language AST, deterministic, zero hallucination
Extraction	Claude (parallel subagents)	Reads anything, outputs structured graph data
Visualization	Obsidian vault	Native graph view, wikilinks, no server needed

No Neo4j required. No dashboards. No server. Runs entirely locally.

Files

graphify/
├── detect.py     detect file types, auto-exclude venvs/caches/node_modules; scan .graphify/memory/
├── extract.py    AST extraction (13 languages via tree-sitter) + call-graph pass (INFERRED edges)
├── build.py      assemble NetworkX graph from extraction JSON; schema-validates before assembly
├── cluster.py    Leiden community detection, cohesion scoring
├── analyze.py    god nodes, bridge nodes, surprising connections, suggested questions, graph diff
├── report.py     render GRAPH_REPORT.md
├── export.py     Obsidian vault, graph.json, graph.html, graph.svg, Neo4j Cypher, Canvas
├── ingest.py     fetch URLs (arXiv, Twitter/X, PDF, any webpage); save Q&A to .graphify/memory/
├── cache.py      SHA256-based per-file extraction cache; check_semantic_cache / save_semantic_cache
├── security.py   URL validation (http/https only), safe fetch with size cap, path guards, label sanitisation
├── validate.py   JSON schema checks on extraction output
├── serve.py      MCP stdio server — query_graph, get_node, get_neighbors, shortest_path, god_nodes
└── watch.py      fs watcher, writes flag file when new files appear

skills/graphify/
└── skill.md      the Claude Code skill — the full pipeline the agent runs step by step

ARCHITECTURE.md   module responsibilities, extraction schema, how to add a language
SECURITY.md       threat model, mitigations, vulnerability reporting
worked/           eval reports from real corpora (httpx, mixed-corpus)
tests/            212 tests, one file per module
pyproject.toml    pip install graphify  |  pip install graphify[mcp,neo4j,pdf,watch]