write-graph-query

name	write-graph-query
description	Write a bounded, typed traversal. Cypher examples first, then ISO GQL, then Gremlin or SPARQL 1.1 stubs. Flag unbounded * and untyped -[]- as bugs. Includes GraphRAG local/global/DRIFT query patterns.

Skill: write-graph-query

Invoked by: graph-query-engineer (primary).

When to invoke: "write this Cypher/GQL"; "why is this path slow?"; "local vs global GraphRAG search."

Output: language pick + bounded pattern + index advice + lint result. Fill ../../templates/query-language-decision.md.

Procedure

1. Pick the language from the engine + model, not the brand. Default teaching order: Cypher first, then GQL equivalent, then Gremlin / SPARQL 1.1 stubs. See ../../knowledge/graph-languages-and-engines-2026.md.
2. Type every relationship. Bound every quantified path (*1..N or GQL {1, N}).
3. If the user pasted -[*]-> or -[]-, rewrite it; do not “just add LIMIT.”
4. If the job is PageRank / community / global shortest-path over a projection, say algorithm library — do not invent CALL algo.*.
5. GraphRAG query section (only after a retrieval graph exists):
   • Local — bounded typed neighborhood of seed entities.
   • Global — read community summaries (not an unbounded walk).
   • DRIFT — local first, escalate to community text. Construction stays on construct-retrieval-graph.
6. Lint snippets with ../../scripts/lint_graph_shape.py. Pre-GQL * in a .gql file is a WARNING.

Guardrails

• Advisory only — the user runs the snippet. No driver.
• Not Neo4j-only. Not GraphQL (.graphql → graphql-engineering).
• Date any engine claim [verify-at-use].
• SPARQL is 1.1 only.