Spec-driven AI Migration: How Nearly 2 Years of Work Was Completed in One Week
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
An Atlassian engineering team migrated 100+ GraphQL fetchers from a legacy gateway to a new platform in one week — work estimated at nearly two years — using a spec-driven AI approach. Each migration unit was captured as a structured behavioral specification in Jira tickets, covering GraphQL schema, upstream APIs, transformations, business rules, and explicit acceptance criteria. AI agents generated migration code from these specs while humans reviewed the specs rather than the generated code line-by-line. Verification was layered: acceptance criteria embedded in tickets, a pre-review validation agent comparing local output against the live service, and a shadow traffic framework for live parity testing. Key lessons include treating the spec as the primary product, using Jira tickets as persistent agent memory to prevent context rot across sessions, and investing equally in verification systems as in code generation.
