Graphify (Knowledge Graph Builder)
Graphify is a community Claude Code skill that turns any input — code, docs, papers, images — into a clustered knowledge graph. Output: HTML visualization + JSON + audit report. Trigger: /graphify <input>.
Graphify is X-ray for codebases — beam in (LLM extraction), develop the film (clustering), see the bones (entity-relation graph). Diagnostic, not surgical: verify the unclear shadows.
Pipeline: parse input → LLM extracts (entity, relation, entity) triples → community detection clusters related entities → renders interactive force-directed HTML viz alongside machine-readable JSON. Use cases: onboarding to unfamiliar codebases (call graph + module map), summarizing research papers (concept map), exploring large doc sets (topic clustering). The audit report flags low-confidence extractions so you know what to verify by hand.
Under the hood: chunks oversized inputs and runs extraction in batches, then merges entities by canonical form (string normalization + LLM-assisted deduplication on close matches). Community detection is typically Louvain or Leiden for modularity-based clustering. The HTML uses D3 or vis.js for the force layout. Limitations: extraction precision drops on jargon-heavy domains without a glossary in the prompt; relation labels are noisy without prompt tuning. Memory cost ≈ O(N) entities; running on a 100K-line codebase produces ~10K nodes and a graph that needs filtering before it's usable.
Graphify is a community skill that turns arbitrary input into a clustered knowledge graph with HTML and JSON output. I use it to onboard onto unfamiliar codebases by visualizing call graphs and module relationships, or to summarize research papers as concept maps. The pipeline is LLM-driven entity and relation extraction, then community detection, then an interactive force-directed visualization. The audit report flags low-confidence extractions so I know where verification is needed before treating the graph as ground truth.
Trusting graphify output without verification on jargon-heavy or domain-specific content. LLM-extracted relations are noisy — always cross-check the audit report's low-confidence list before citing the graph as evidence.