Graphify — Folder-to-Knowledge-Graph
Turns any folder of code, docs, papers, images, or video into a queryable knowledge graph so AI can answer real questions about your project.
Some data-flow disclosure: claims to read only GEMINI/GOOGLE keys, skips sensitive files, local Whisper prompts. But SKILL.md instructs the agent to auto-install the third-party PyPI package 'graphifyy' (with a --break-system-packages fallback), an unverified supply-chain dependency; it auto-clones arbitrary GitHub repos into ~/.graphify; no rollback instructions beyond graphify-out. Deducted for: unconfirmed external installs and arbitrary cloning, more-than-least-privilege behavior, incomplete recovery.
Instructions are structurally coherent with caching/chunking and partial failure handling (zero files, transcription failure, cache hits), but there is zero test evidence; key paths (AST extraction, semantic subagents, Neo4j push) are unverifiable statically; SKILL.md is truncated mid-specification at the node-ID rules, so later steps are missing and execution could stall. Deducted for: no tests, truncated file, thin diagnosable failure feedback.
Scenario is clear (knowledge-graph queries over code/docs) with detailed parameters and outputs; however the description's trigger is over-broad ('any question about a codebase', taking over whenever graphify-out/ exists), risking false invocation; core function depends on GitHub, PyPI, Gemini/Whisper with no declared Chinese support and poor mainland-China reachability. Deducted for: loose trigger boundaries, undeclared non-fit ranges, environment reachability risk.
Repo is MIT-licensed, SKILL.md is layered (usage/purpose/steps), README shows maintainers and contribution paths; but the skill itself has no version, changelog, or known-limitations section, the publisher is registry-unverified, and the truncation suggests incomplete maintenance. Deducted for: missing versioning/governance and undisclosed limits.
Claimed outputs (HTML graph, GraphRAG JSON, GRAPH_REPORT.md) would beat manual graph-building, but static review cannot confirm any output is directly usable; entries like '--html is a no-op' hint at doc/behavior drift. Deducted for: no representative output evidence; comparative benefit rests on claims.
Only one skill instruction file; no tests or CI coverage for the skill (repo CI covers frontend builds and deploy tokens only), no third-party execution evidence. Deducted for: no independently reproducible material behind key claims.
- The skill auto-installs the third-party PyPI package 'graphifyy' (including a --break-system-packages fallback) — a supply-chain trust risk; audit the package before first run.
- The description's trigger is overly broad: it takes over any codebase question whenever graphify-out/ exists, risking false invocation; prefer explicit /graphify invocation only.
- Core function depends on GitHub, PyPI and Gemini APIs; mainland-China reachability is poor and Chinese-language support is undeclared.
- SKILL.md is truncated at the node-ID specification; runtime may stall on missing instructions.
- The skill has no version or changelog; update behavior is unpredictable.
What does this skill do, and when should you use it?
Graphify is a Claude Code skill invoked with /graphify that extracts an entity-relationship knowledge graph from a folder or GitHub repository. It detects files, optionally transcribes audio/video, then combines deterministic AST extraction of code with LLM subagent (or Gemini API) semantic extraction, followed by community detection. It produces three outputs — interactive HTML, GraphRAG-ready JSON, and a plain-language GRAPH_REPORT.md — and persists the graph in graphify-out/ so later questions query it directly. Every edge carries an honest EXTRACTED/INFERRED/AMBIGUOUS audit label, and multiple repos can be merged into one cross-repo graph.
1) Detects code, docs, PDFs, images, and audio/video in a target path, warning on oversized corpora; 2) transcribes video/audio to text via Whisper with a domain-hinted prompt; 3) runs free deterministic AST extraction on code in parallel with LLM subagents over docs/images (cached, chunked 20-25 files each); 4) merges extraction, builds the graph, runs community detection, finds god nodes and surprising cross-document connections; 5) generates GRAPH_REPORT.md, graph., interactive HTML, plus optional Obsidian vault, SVG, GraphML, Neo4j Cypher export/push, and an agent-crawlable wiki; 6) offers /graphify query with BFS/DFS traversal to answer natural-language questions, /graphify path for shortest paths between concepts, and an MCP stdio server exposing query_graph and related tools to other agents.
- A developer onboarding onto an unfamiliar codebase: clone the repo, run /graphify, then ask 'How does X work?' or 'What calls Y?' without rereading all sources
- A research group: drop a folder of papers and notes in and let community detection surface cross-document concept links
- A multi-service team: extract each subfolder of a monorepo separately, then merge into one cross-service graph with merge-graphs
- A content creator or podcaster: transcribe recorded audio/video and graph it together with written notes into an Obsidian vault
- A Neo4j user: push exported Cypher into an existing Neo4j instance to plug into current graph workflows
What are this skill's strengths and limitations?
- Extremely broad input: code, docs, PDFs, images, audio, and video all become graph nodes
- AST structural extraction is free and deterministic; semantic results are cached, keeping reruns cheap
- EXTRACTED/INFERRED/AMBIGUOUS audit trail makes graph confidence transparent
- Rich export ecosystem: HTML, GraphRAG JSON, Obsidian, Neo4j, SVG, GraphML, wiki, MCP server
- Supports incremental updates, cross-repo merging, and a --watch mode that rebuilds on code changes
- Core semantic extraction orchestration depends on Claude Code's parallel subagent mechanism; porting elsewhere requires rework
- Without a Gemini key it leans on the host LLM session, so token cost scales with corpus size
- SKILL.md was presented truncated; full content and standalone test coverage cannot be verified
- It sits inside a career-platform (Career Pilot) repository, so its update cadence is tied to an unrelated project
- Corpora over ~2M words or 500 files require narrowing to a subfolder before extraction
How do you install this skill?
The skill lives at .claude/skills/graphify/SKILL.md in the career-pilot repo; place it in your Claude Code skills directory (typically your project's .claude/skills/). On first run it auto-installs the Python package: uv tool install --upgrade graphifyy, or pip install graphifyy without uv. Optional: set GEMINI_API_KEY or GOOGLE_API_KEY for Gemini-based semantic extraction (pip install 'graphifyy[gemini]'). The repo README describes the Career Pilot web application, not this skill; treat SKILL.md as the authoritative install reference.
How do you use this skill?
Trigger /graphify in Claude Code. Common forms: /graphify <path> builds a full graph; /graphify https://github.com/<owner>/<repo> clones then builds; --update for incremental re-extraction; --mode deep for richer INFERRED edges; --neo4j-push bolt://localhost:7687 to push to Neo4j; --mcp to start an MCP server; /graphify query "<question>" to query an existing graph (add --dfs or --budget N as needed); /graphify explain "<node>" for plain-language explanations; /graphify --help prints usage. Once graphify-out/graph. exists, natural-language codebase questions automatically take the fast path and query the graph without rebuilding.
How does this skill compare with similar options?
The skill positions itself against generic GraphRAG pipelines (its JSON is 'GraphRAG-ready') and integrates with existing graph tooling like Neo4j, Gephi/GraphML, and Obsidian, but the source material names no direct one-shot competitor.