Netmarble Nexus Establishes AI Coding Culture in Game Development Organization with Kiro Adoption

Netmarble Nexus successfully adopted Kiro in its game development organization, increasing the utilization of AI coding tools. Initially, barriers included AI limitations with complex game code and negative perceptions, but these were overcome through a practitioner-centric Champion system and executive support.

Internal Adoption Strategy

The core of the adoption was selecting internal practitioners as Champions rather than external experts, enabling them to create Steering files based on domain knowledge. A pilot was applied starting with new development projects to reduce risk, and rules created by Champions were shared across the team via version control, driving voluntary participation. Through this, developers experienced that AI could provide accurate responses tailored to project context, beyond merely generating code.

Steering File Structure and Automation

Steering files are a key mechanism for conveying project rules and background knowledge to AI. Netmarble Nexus segmented these into main guidelines, architecture, and domain documents to implement efficient context management. In particular, the 'prohibition + counter-example' format was used to prevent common AI mistakes, and rule documents were continuously evolved through Hooks that automatically log and refine rules upon completion of agent tasks.

MCP Integration and Expanded Use Cases

Through MCP (Model Context Protocol) integration, AI gained direct access to databases and Excel files. This enabled automation of repetitive tasks such as verifying data consistency in natural language without writing SQL, or automatically detecting discrepancies between log definitions and server code. However, strict security prerequisites were maintained, such as limiting connections to development environments rather than live service data.

Results and Challenges

As a result of adoption, legacy code analysis time was reduced, and knowledge levels within teams were leveled up. However, review burden due to increased code volume generated by AI and difficulties in introducing unit tests for legacy code remain unresolved challenges. Additionally, bottlenecks occurred at the planning stage as development speed improved, confirming the need for balanced improvements across the entire development process.