A Claude Code skill that teaches an agent to design the shape of the work before doing the work. Instead of running a multi-step task as a straight line, the agent maps it as a graph: nodes are units of work, edges are data dependencies. Independent work fans out in parallel, findings pass through fresh-context verifiers, and parallel writers get isolated workspaces.
The method comes from the "graph engineering" idea: the successor to loop engineering. A single agent improving one thing on repeat is a loop. A network of agents that watch and correct each other is a graph. Read the original article by @0xCodila for the full story.
When the agent plans any multi-step task (a codebase audit, a migration, a review, a research sweep), the skill makes it:
- Find the real edges. For every "and then," ask whether the next step actually reads the previous step's output. If not, the wait is wasted.
- Fan out the independent work, one agent per file, source, or angle.
- Verify findings on independent context. A fresh verifier node checks real signals, never the executor's own claims.
- Isolate before parallelizing. Decide where each agent works, how results merge, and what happens when two disagree.
It also tells the agent when a graph is the wrong tool: small tasks, exploratory work, tasks that need step-by-step oversight, and true chains where every step reads the last one's output.
Copy the skill folder into your personal Claude Code skills directory:
git clone https://github.com/mlopscommunity/graph-engineering-skill.git
cp -r graph-engineering-skill/graph-engineering ~/.claude/skills/Claude Code picks it up automatically. It triggers when you plan or build multi-step work, or you can invoke it directly with /graph-engineering.
graph-engineering/SKILL.md: the skill itself.
MIT