graph-engineering

# Graph Engineering Engineer the graph. Do not run it. Design the work around AI instead of relying on one giant chat. ## Design-only boundary The sole deliverable of this skill is an approval-ready graph. This boundary applies in every mode, including Plan Mode. Treat verbs such as *determine*, *research*, *validate*, *build*, or *implement* as the outcome the future graph must produce—not as permission to perform that work now. While designing: - inspect the user's context only enough to understand the objective, constraints, and consequential decisions; - ask only questions whose answers could materially change the graph; - identify the evidence and access needed to resolve those decisions; - turn researchable unknowns into jobs inside the graph; - place a human gate before any future evidence method that needs new authority; - do not browse or investigate those unknowns; - do not launch workers, subagents, skeptics, or mergers; - do not execute any graph job or produce the graph's terminal deliverable; and - do not modify the target project or external systems. Stop after presenting the graph for human approval. The displayed nodes are future jobs. Execution begins only after the user has seen and approved the graph, and it happens outside this skill. ## Align before engineering Begin every invocation with an adaptive alignment interview. First inspect the conversation, supplied artifacts, current authoritative context, and previously settled decisions. Do not ...