The Problem

Building knowledge graphs for AI agents is usually ad hoc — teams bolt together extraction scripts and dump triples into a store without a schema, verification, or provenance. The result is a pile of facts that agents can't reliably query. This repo packages a disciplined 9-stage pipeline (from SEU's graduate course) plus task-graph orchestration patterns into a Claude skill, so an agent can build or teach the methodology on demand.

What This Does

The repo is a Claude skill — a directory of Markdown files that teach an agent how to design and build knowledge graphs and task graphs. It's not a runtime library; it's structured knowledge. The core is graph-engineering/SKILL.md, which instructs the agent on the 9-stage pipeline: scope → representation → ontology → entities → relations → events → quality gate → fusion → serve to LLMs. The references/ folder holds the distilled course material (curriculum.md, modeling.md, extraction.md, fusion-and-llm.md, task-graphs.md). WORKFLOWS.md provides nine paste-ready prompt blocks (/kg-tutor, /kg-scope → /kg-rag) that chain into a full build workflow.

How It Is Wired

This is a documentation repo, not an application. There is no executable code, no call graph, and no entry point that runs anything. The "wiring" is the skill's activation path: an agent reads graph-engineering/SKILL.md and follows its instructions. The dist/graph-engineering.skill file is the packaged skill bundle. The only external dependency is the original Chinese course repo (npubird/KnowledgeGraphCourse), which is referenced but not redistributed.

Each file's responsibility:

  • graph-engineering/SKILL.md — the agent-facing instructions; the primary entry point.
  • graph-engineering/references/*.md — the actual course content, organized by pipeline stage.
  • WORKFLOWS.md — copy-paste prompt templates for interactive use.
  • dist/graph-engineering.skill — the packaged skill for distribution.
  • README.md / LICENSE — usage and MIT licensing.

How To Use It

Setup — clone and copy the skill into Claude's skills directory:

git clone https://github.com/moses-y/graph-engineering.git
cp -r graph-engineering/graph-engineering ~/.claude/skills/

Running it — no code to execute. After installation, prompt your agent with either "build a knowledge graph from my docs" or "teach me graph engineering." The skill's teaching mode walks through the pipeline stage by stage, using your project as the running example.

Configuration — none. No environment variables, secrets, or config files required.

Real-World Use

A practical scenario: a team has a pile of internal documentation and wants a queryable knowledge graph. They install the skill, then prompt: "Build a knowledge graph from ./docs/ following the 9-stage pipeline." The agent reads SKILL.md, applies the ontology-first approach, models the domain before extraction, and produces a schema-verified graph with provenance on each edge. The task-graph patterns in references/task-graphs.md then help orchestrate the extraction work itself — parallel workers, separate verifier contexts, and a human gate at the fusion step.

Code Health & Issues

Static analysis (heuristic, verify against the repo) reports:

  • Med/SDLC — No test files detected — untested code paths — repository-wide. Expected for a documentation-only repo; nothing to execute or test.
  • Med/SDLC — No CI/CD pipeline detected — no automated build/test gate — .github/ or CI config. Again, expected for Markdown-only content.

No committed secrets, no lockfile, no Dockerfile. MIT license present. The absence of tests and CI is not a real risk here — there is no code to break — but it does mean changes to the skill's instructions are unverified until an agent runs them.

The Bottom Line

This is a well-structured, opinionated knowledge pack for teams that want a disciplined approach to knowledge-graph construction. It's not a tool — it's a methodology encoded for AI agents. The task-graph material (stop rule, diamond pattern, human gate) is genuinely useful and grounded in published research. Use it if you want your agent to build graphs with a schema and verification, not just dump triples. If you're looking for executable code, this isn't it.