graphify-knowledge-graph

---
name: graphify-knowledge-graph
description: Build queryable knowledge graphs from code, docs, papers, and images using AI coding assistant skills
triggers:
  - "graphify my codebase"
  - "build a knowledge graph"
  - "turn my files into a graph"
  - "understand this codebase with graphify"
  - "run graphify on this folder"
  - "query the knowledge graph"
  - "install graphify skill"
  - "extract relationships from my code"
---

# graphify-knowledge-graph

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

graphify turns any folder of code, docs, papers, or images into a queryable knowledge graph. It runs as an AI coding assistant skill — type `/graphify` in Claude Code, Codex, OpenCode, or OpenClaw to extract structure, relationships, and design rationale from your files into an interactive graph you can navigate and query without re-reading raw files.

---

## Install

```bash
pip install graphifyy && graphify install

The PyPI package is graphifyy; the CLI and skill command remain graphify.

Platform-specific install

graphify install                        # Claude Code (default)
graphify install --platform codex       # Codex
graphify install --platform opencode    # OpenCode
graphify install --platform claw        # OpenClaw

Always-on assistant integration (recommended)

Run once per project so your assistant consults the graph before searching files:

graphify claude install      # writes CLAUDE.md section + PreToolUse hook (Claude Code)
graphify codex install       # writes AGENTS.md (Codex)
graphify opencode install    # writes AGENTS.md (OpenCode)
graphify claw install        # writes AGENTS.md (OpenClaw)

Undo with the matching uninstall command:

graphify claude uninstall

Manual install (curl, no pip)

mkdir -p ~/.claude/skills/graphify
curl -fsSL https://raw.githubusercontent.com/safishamsi/graphify/v3/graphify/skill.md \
  > ~/.claude/skills/graphify/SKILL.md

Add to ~/.claude/CLAUDE.md:

- **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.

Core workflow

1. Build the graph

# In your AI coding assistant
/graphify .                        # current directory
/graphify ./src                    # specific folder
/graphify ./raw --mode deep        # aggressive INFERRED edge extraction
/graphify ./raw --no-viz           # skip HTML, produce report + JSON only

2. Outputs

graphify-out/
├── graph.html       # interactive — click nodes, search, filter by community
├── GRAPH_REPORT.md  # god nodes, surprising connections, suggested questions
├── graph.json       # persistent graph — query later without re-reading files
└── cache/           # SHA256 cache — re-runs only process changed files

3. Query the graph

/graphify query "what connects attention to the optimizer?"
/graphify query "what connects attention to the optimizer?" --dfs        # trace a path
/graphify query "what connects attention to the optimizer?" --budget 1500  # cap tokens
/graphify path "DigestAuth" "Response"      # shortest path between two nodes
/graphify explain "SwinTransformer"         # expand a single node

Key commands reference

Building

Ingesting remote content

/graphify add https://arxiv.org/abs/1706.03762          # fetch a paper, add to graph
/graphify add https://x.com/karpathy/status/...         # fetch a tweet
/graphify add https://... --author "Andrej Karpathy"    # tag original author
/graphify add https://... --contributor "Your Name"     # tag who added it

Querying

/graphify query "why does the auth layer depend on redis?"
/graphify query "what implements the retry protocol?" --dfs
/graphify path "Transformer" "AdamW"
/graphify explain "DigestAuth"

Exporting

/graphify ./folder --svg              # export graph.svg
/graphify ./folder --graphml          # export graph.graphml (Gephi, yEd)
/graphify ./folder --neo4j            # generate cypher.txt for Neo4j import
/graphify ./folder --neo4j-push bolt://localhost:7687   # push to live Neo4j
/graphify ./folder --obsidian         # generate Obsidian vault (opt-in)
/graphify ./folder --wiki             # build agent-crawlable wiki (index.md + per-community articles)
/graphify ./folder --mcp              # start MCP stdio server

Git hooks

graphify hook install     # post-commit + post-checkout: auto-rebuild on commit/branch switch
graphify hook uninstall
graphify hook status

Supported file types

How it works

graphify runs in two passes:

1. AST pass (deterministic, no LLM): Extracts classes, functions, imports, call graphs, docstrings, and rationale comments from code files.
2. Semantic pass (parallel Claude subagents): Extracts concepts, relationships, and design rationale from docs, papers, and images.

Results are merged into a NetworkX graph, clustered with Leiden community detection (topology-based — no embeddings or vector database), and exported as HTML, JSON, and a plain-language audit report.

Edge provenance tags

Every relationship is tagged so you always know what was found vs guessed:

Python API

graphify is primarily a CLI/skill tool, but the graph output (graph.json) is standard NetworkX JSON you can load and traverse:

import json
import networkx as nx

# Load the persistent graph
with open("graphify-out/graph.json") as f:
    data = json.load(f)

G = nx.node_link_graph(data)

# Find god nodes (highest degree)
god_nodes = sorted(G.degree(), key=lambda x: x[1], reverse=True)[:10]
for node, degree in god_nodes:
    print(f"{node}: {degree} connections")

# Find all EXTRACTED edges (high confidence, found in source)
extracted_edges = [
    (u, v, d) for u, v, d in G.edges(data=True)
    if d.get("provenance") == "EXTRACTED"
]

# Find INFERRED edges above a confidence threshold
high_confidence_inferred = [
    (u, v, d) for u, v, d in G.edges(data=True)
    if d.get("provenance") == "INFERRED" and d.get("confidence_score", 0) > 0.85
]

# Shortest path between two concepts
try:
    path = nx.shortest_path(G, source="DigestAuth", target="Response")
    print(" -> ".join(path))
except nx.NetworkXNoPath:
    print("No path found")

# Get all nodes in a community
communities = {}
for node, data in G.nodes(data=True):
    community_id = data.get("community")
    if community_id is not None:
        communities.setdefault(community_id, []).append(node)

for cid, members in sorted(communities.items()):
    print(f"Community {cid}: {', '.join(members[:5])}{'...' if len(members) > 5 else ''}")

Working with rationale nodes

graphify extracts # NOTE:, # IMPORTANT:, # HACK:, # WHY: comments and docstrings as rationale_for nodes:

# Find all rationale nodes and what they explain
rationale_nodes = [
    (node, data) for node, data in G.nodes(data=True)
    if data.get("node_type") == "rationale_for"
]

for node, data in rationale_nodes:
    print(f"Rationale: {data.get('label')}")
    # Find what this rationale is connected to
    neighbors = list(G.neighbors(node))
    print(f"  Explains: {neighbors}")

Querying semantic similarity edges

# Find cross-file semantic links (concepts connected without structural relationship)
semantic_edges = [
    (u, v, d) for u, v, d in G.edges(data=True)
    if d.get("relation") == "semantically_similar_to"
]

for u, v, data in semantic_edges:
    score = data.get("confidence_score", 0)
    print(f"{u} ~ {v} (confidence: {score:.2f})")

Common patterns

Pattern 1: Onboard to an unfamiliar codebase

# Install graphify, build the graph, read the report
pip install graphifyy && graphify install

# In Claude Code
/graphify .

# Read the output — god nodes tell you what everything routes through
cat graphify-out/GRAPH_REPORT.md

Pattern 2: Mixed research corpus (Karpathy-style /raw folder)

Drop code, PDFs, screenshots, and notes in one folder:

raw/
├── attention_is_all_you_need.pdf
├── training_notes.md
├── whiteboard_photo.png
├── nanoGPT/
│   └── model.py
└── tweet_screenshot.jpg

/graphify ./raw

graphify uses Claude vision on images, citation mining on PDFs, AST on code, and semantic extraction on markdown — all merged into one graph.

Pattern 3: Incremental updates (large codebase)

# First full build
/graphify ./src

# After making changes — only re-processes changed files via SHA256 cache
/graphify ./src --update

# After a major refactor — rerun clustering without re-extracting
/graphify ./src --cluster-only

Pa

Content truncated.